Circle CEO Jeremy Allaire: When AI Meets Blockchain, the Familiar 'Company' is About to Disintegrate

Original Author: Jeremy Allaire, Co-founder and CEO of Circle

Compiled by: Jiahua, ChainCatcher

I. Technological Integration and the Disintegration of Companies

Every platform-level transformation in the internet era is not driven by a single invention, but rather by several mature technologies colliding at a certain point in time. The birth of the Web required graphical interfaces, a commercially open internet, sufficiently fast modems, along with a set of open software layers including web pages, links, and servers, all of which are indispensable.

Digital media, mobile internet, cloud computing, and social platforms all follow the same path. Behind this is a recurring pattern: when multiple capabilities converge, the marginal cost of certain originally expensive activities collapses to near zero; once costs collapse, the speed of these activities explodes. The Web ignited the speed of information dissemination, mobile and social ignited the speed of interpersonal communication, and cloud computing ignited the speed of software production and delivery.

Now, two new "operating systems" are merging, applying the same mechanisms to two things that have never been natively digitized on the internet: intelligence and economic activities themselves.

The first is the operating system of intelligence, which takes the form of artificial intelligence built on foundational models and intelligent agent systems. The second is the operating system of the economy, which is the blockchain network where value, contracts, and collaboration can be expressed and executed through software. The former drives the cost of cognition and work towards zero, while the latter drives the cost of transactions, settlements, and collaborations towards zero.

The two reinforce each other. Intelligence allows economic activities to operate at machine speed, while the economic foundation enables machine intelligence to trade, exchange value, collaborate, and execute contracts. The core assertion is that the intelligent economy and the on-chain economy are not neighbors but rather part of the same economic entity, and both are converging into a force that reshapes the global economic system.

First, let's look at the operating system of intelligence. It consists of the capabilities of cutting-edge foundational models and the reasoning and infrastructure that enable the models to scale and execute work. Today's representatives include platforms like Claude, Claude Code, OpenAI, and Codex.

This is a new type of computing machine: instead of programming in the traditional way, it receives instructions in natural language to produce results and complete tasks. The atomic unit of this work is the intelligent agent, which is the reasoning process sent to execute a specific task.

Why is this important? First, we need to understand what a company really is.

Stripped of brands and buildings, a company is essentially an information system organized around a set of familiar functions: product and engineering, marketing, sales, talent, finance, legal compliance, operations, and customer service. The cost of maintaining this system largely comes from human resources.

Looking at the entire economy, human resources are the largest single operational expense, typically accounting for a quarter to a third of revenue, with even higher ratios in the service industry. In knowledge-based and technology companies, this is nearly absolute: non-capital expenditures are almost entirely wages. In other words, such companies are essentially "organized cognition with a logo."

Outside the company walls lies a second massive market: professional services such as consulting, legal, accounting, and agency services, which are also organized human resources rented from outside. These two enormous cost pools are precisely what the intelligent operating system targets.

This is why the intelligent economy disrupts classical corporate theory. Economists have long used transaction costs to explain why companies exist: the costs of coordinating, contracting, and trusting external labor are too high, so companies internalize work that is "cheaper to do themselves." The boundaries of a company are essentially defined by coordination costs.

When every non-physical work unit can be completed by a discoverable, contractable, and instantly settleable intelligent agent, coordination costs begin to collapse, and the traditional boundaries of companies lose their meaning.

The most intuitive result is the one-person company: one person directs a group of intelligent agents to complete work that previously required multiple departments. Within large companies, small teams with super high leverage will emerge, executing business at a scale far exceeding their own staffing.

The economic calculus continues to compound, as three exponential curves are moving simultaneously: cognitive work is increasingly transferred to intelligent agents, the proportion of human resources in operational costs is declining; the cost of running intelligent agents is continuously decreasing, with the price of equivalent machine intelligence dropping by an order of magnitude each year; meanwhile, the capabilities of intelligence are consistently improving across almost all benchmarks.

Cheaper, more powerful, and bearing more costs, the multiplication of these three factors will unleash tremendous productive potential.

This disintegration will not happen uniformly. It will first appear in the field of software engineering, as today's models are exceptionally good at understanding and writing code. At the same time, it will parallelly unfold in marketing, sales, customer service, and a large number of legal, financial, and compliance tasks. Any work involving refining, launching, analyzing, and presenting information falls into this category.

Physical labor is the furthest from this change. The main contribution of robots in heavy industry and assembly is still enhancement rather than complete replacement, and the challenges in the physical production sector may take over a decade to resolve.

However, this disintegration should not be simply understood as "reducing personnel." A more accurate picture is one of coexistence between enhancement and replacement.

Human creativity is further amplified with the help of deep intelligent agent skills, allowing individuals to take on broader cross-functional roles and switch work focuses at unprecedented speeds. Some capabilities still belong irreplaceably to humans, such as emotional relationships and face-to-face work, critical judgment of intelligent agent processes, and governance and accountability responsibilities that cannot be delegated to machines.

Here lies a clear contradiction: for individuals, AI agents are amplifying human capabilities; but from the perspective of the overall economy, as machines take on more work, how much of the new output will flow to humans in the form of wages and labor income?

II. Assembly, Coordination, and Why Companies Need to Go On-Chain

After companies disintegrate into intelligent agent skills, the question is no longer "what can be automated," but rather how these fragments can be reassembled into collaborative work.

The mechanism to achieve this is a choreography layer built on advanced foundational models. At its core is a master choreography agent: faced with any task, it breaks down the goal into a task pipeline and assigns tasks to different sub-agents. Surrounding infrastructure is responsible for initializing the pipeline, maintaining context and memory, executing tasks, and recombining returned results.

The same universal architecture can serve any function. Marketing pipelines, financial pipelines, product pipelines, or sales pipelines are structurally the same machine aimed at different tasks.

Humans are not absent; they occupy two positions.

Some people are "in the loop," completing or reviewing specialized work that requires human judgment within the pipeline; others are "on the loop," responsible for setting goals, defining acceptance criteria, supervising output quality, and deciding when machines should stop to ask humans questions.

This is the specific form of human oversight in intelligent agent companies. The infrastructure supporting it is gradually becoming widespread, and many leading teams have begun to fully implement this new architecture.

The key is that this begins within companies but will not remain confined to them.

To orchestrate their work, companies must transform each function into a clearly defined skill: train in specific areas, access the right data, and continuously update. But a skill clear enough to be orchestrated internally in a company is naturally also clear enough to be discovered and hired externally.

Once internal modularization meets the intelligent agent market, intelligent agent payments, and smart contracts, the disassembly that companies complete for optimization will become the foundation for cross-organizational markets.

The open intelligent agent economy will become a byproduct of companies optimizing themselves, without anyone needing to deliberately create it.

This market may evolve into two forms.

The first is enterprises purchasing intelligence by usage from a few large platforms; the second is forming a true intelligent agent labor market, where companies hire specialized intelligent agents to complete specific tasks.

The second form is more likely to emerge and is more important, for the same reasons that have long been observed in the software industry: deep domain knowledge has lasting value.

Foundational models will gradually become commoditized inputs, and the truly lasting businesses will be those specialized intelligent agents deeply engaged in specific fields, such as creative marketing, video production, intellectual property, contract negotiation, and thousands of specialized crafts.

They build barriers by aggregating proprietary contexts and specialized data, continuously refining capabilities, and achieving enterprise-level security and reliability standards. Specialized intelligent agents can be discovered through registries and markets, with their metadata being readable by humans and directly callable by other intelligent agents, leading to fierce competition among them.

Another economic detail worth noting is that due to fierce competition among foundational models, specialized intelligent agents will route tasks among multiple models to optimize their intelligence costs.

Thus, models become cost items, while intelligent agents become the business itself.

However, the intelligent agent labor market will soon encounter a hard problem, and solving this problem is precisely why the entire system must go on-chain.

Before hiring an intelligent agent, orchestrators must know that this agent is real, its work is trustworthy, and there is accountability in case of issues. When a "worker" is a piece of software that could be assembled anywhere in the world, these conditions do not naturally hold.

The solution is that the identity of an intelligent agent is not a single entity but rather a multi-layered structure.

The bottom layer is cryptographic verifiability. An economic operating system built on public chains allows data, transactions, and code executions to be verified in real time, with the foundation of trust being cryptography rather than any intermediary.

However, minimizing trust only applies to things that the system can self-verify: whether a transaction occurred, whether a balance changed, whether a contract executed according to code.

It cannot adjudicate facts about the external world, resolve disputes, or reverse a "code correct, reality wrong" outcome. These issues must be handled by accountable external mechanisms, including oracles that prove external facts, arbitration mechanisms for dispute resolution, and human intervention when necessary.

Thus, the overall architecture is formed: the core ensures integrity, while the edges ensure accountability.

On this foundation, several additional layers are needed to ensure that intelligent agents truly have "someone responsible."

The first layer is anchored in the real world. The work of agents ultimately needs to be tied to a real, verified entity. Financial infrastructure companies have large-scale operational compliance identity verification systems that can answer several key questions: Who created this agent? Is the creator legitimate? Is their reputation good?

The second layer is the economic existence of the agent itself, including the wallets it controls and verifiable credentials with real-world associated information.

Above that is reputation. Reputation accumulates over time through work records and user evaluations, and because it is anchored to a verified real identity, it is more resistant to fraud than a one-time pseudonym.

This is also why companies need to be on-chain, rather than simply trusting a private database of a market platform.

Private registries bind trust to a single operator; on-chain systems, however, use cryptography and real identity anchoring to make trust portable, allowing it to flow across markets, companies, and borders without needing to trust the owners of any specific platform.

What a global trading open agent economy needs is precisely this capability, which no private database can provide.

These layers together form a chain of accountability: every action of the agent can be traced back through wallets and credentials to a verified, reputable real creator.

In this economic system, autonomy does not equal anonymity. Autonomous agents must be accountable agents.

This chain allows counterparties to confidently hire a piece of software, enables regulators to find responsible parties, and ensures that machine autonomy does not slip into unaccountable actions.

At this point, the restructured company has taken shape: a small core of humans is at the center, responsible for setting goals and exercising judgment; orchestrators coordinate a workflow composed of specialized agents, some built by the company and more hired from the global market; each collaboration is a contract executed by software that can be enforced; every actor, no matter how autonomous, ultimately can be traced back to a responsible person through the accountability chain.

Coordination is no longer just an internal management issue but has become an economic issue completed across company boundaries in software.

The agent company thus reveals another side of itself: the on-chain company.

But all of this presupposes a premise that has yet to be established: the need for a type of currency that agents can hold, exchange at machine speed, and circulate in large volumes and small increments, while not having to bear the risk of the currency itself in every transaction.

III. Monetary Foundation: Speed, Security, and Finality

The company restructured in the previous section still lacks a key element: a currency that agents can hold and exchange at machine speed.

They need to transfer currency in huge amounts and tiny increments without having to reassess the reliability of the money itself in every transaction. This last condition is crucial.

The attributes that allow currency to be used by software at extremely high speeds are precisely not those possessed by traditional bank currencies. Following this logic, a concrete and "old-fashioned" answer emerges: a fully reserved currency operating on an open network with settlement finality.

Starting with speed, because speed will reorganize everything else.

When the marginal cost of storing and transferring currency approaches zero, transfer times drop to hundreds of milliseconds, and the currency itself can be directly controlled and programmed by software, a currency base with extremely high circulation speed will form.

The same dollar can be put to multiple uses in a short time, regardless of the amount, and can be available for use immediately. Micro-value exchanges orchestrated by agents also become feasible for the first time.

This is merely the unit economics that information and software have already followed on previous internet platforms, now applied to the currency itself.

This often invites opposition from monetary economists: modern banks rely on leverage to create circulation speed, repeatedly lending the same deposit to create risky synthetic dollars. If full reserves prohibit this multiplier, will the economy lack credit?

The answer is no.

The reuse benefits brought by leverage can be obtained without permanently creating risky synthetic dollars. When currency circulates quickly enough, a dollar can be locked for a few seconds and then lent to a third party; the speed itself can replace the multiplier.

Fully reserved currency does not mean funds are idle. Its reserves can be invested in short-term government bonds to finance government spending, and money can still be at work even when it is "static."

Credit will not disappear; rather, it may become stronger. On-chain currency markets can support machine-mediated credit, adjudicated over very short terms, pooling and diversifying funds while allowing lenders to retain instant redemption capabilities similar to demand deposits.

Credit will not be starved by fully reserved currency; instead, it will be rebuilt upon it and become stronger and safer.

Why can't base currency embed any risk? Because the faster the circulation speed, the more dangerous risky currency becomes.

Bank runs are now much faster than in the past. In the mobile banking era, a run can collapse a large institution within hours; in the machine speed era, a run could happen in an instant.

An agent deciding whether to accept a unit of currency needs a one-to-one redeemability guarantee that it never has to doubt. If it has to worry about redeemability, it must price that risk into every transaction, which is economically infeasible in microtransactions at a million times the speed.

There is also a more insidious problem. In a world where thousands of banks issue currency separately, each issuer's dollar is its own IOU, carrying different risks. Dollars from different issuers are not completely equivalent, and prices may deviate.

This would destroy what monetary economics calls "monetary unity": one dollar is one dollar, homogeneous and at par. This is the premise for currency to serve as a large-scale unit of account.

The agent economy is global, operating at internet scale. An agent far away on the other side of the world cannot stop to assess the credit of a strange issuer when making real-time settlement decisions. National-level backstop mechanisms that allow bank currency to function, such as deposit insurance and lender of last resort, cannot cover most participants in a borderless system.

Fully reserved currency is the only form of currency that can maintain par value for everyone, everywhere, without relying on these backstop mechanisms. This is the traditional idea of "narrow banking": 100% reserves. It has been proposed for a long time and has been set aside for a long time because, while it is safe, it is not useful enough.

What changes this conclusion is machine mediation and internet-scale utility. For the first time, they make narrow banking maximally useful. The security of currency units is a necessary condition, but not a sufficient one; settlement must also be equally unquestionable.

Sound currency has always been built on a foundation of "no questions asked": money is money precisely because no one needs to conduct due diligence on it before receiving it. The financial system has written this into the "Principles for Financial Market Infrastructures," whose core guarantee is: when a payment system determines that a transaction has been finally settled, it must truly become the final result.

This guarantee becomes crucial at high speeds, and this is precisely the shortcoming of decentralized networks in the past. Hard forks may no longer recognize settled transactions, and chain reorganizations may roll back transactions. The best promise these systems can offer is often probabilistic finality: after waiting for several confirmations, reversals become less likely.

Agents trading at extremely high speeds cannot be built on "probably final." They need deterministic sub-second finality: once settlement is complete, it takes effect immediately and cannot be changed. This is a specific technical requirement. Chains designed for this purpose can meet it, while old designs struggle to do so.

The trinity of par value, redeemability, and finality is what allows currency to be accepted by machines without inspection.

Settlement finality also brings a superficial paradox: people simultaneously want payments to be reversible, able to issue refunds, prevent fraud, or reverse erroneous payments.

The solution is architectural, just as the internet overlays reliable protocols on a simple, unreliable foundation: keep the base currency with deterministic finality, and then build reversibility as an optional protocol on top, such as custodianship triggered by time and events, refund pools, and insurance for those funds.

Reversibility cannot be directly welded into the currency itself. Doing so would destroy the "no questions asked" attribute of currency, forcing every agent to price rollback risk into every transaction, thus repeating the mistakes of IOU currency.

Pushing reversibility to the edge, making it a composable protocol layer, can also provide protection without compromising the core.

From this perspective, irreversibility is not a hazard to be mitigated but a reliable feature. The reason a final settlement can serve as the foundation for agents to continue building upward is precisely because it never needs to be reconsidered.

These security features will not be automatically realized but require institutional architecture to support them. Currently, this architecture is gradually taking shape.

According to recent legislation, large stablecoin issuers will be supervised by federal banking regulators; stablecoins will adopt bankruptcy isolation structures to separate funds from the collapse of the issuer or partner banks; national trust bank licenses will provide a fiduciary system, separating the base layer currency obligations and credit risks. This is equivalent to the rebirth of narrow banking.

Reserve designs will also become safer as the system expands. Qualified reserves may gradually shift from short-term government bonds to central bank cash and overnight repurchase agreements related to the central bank.

Policymakers are increasingly supporting this direction, and similar trends are emerging in other regions. Proposals from the Bank of England point to the same goal, and European payment system legislation has begun considering allowing electronic money institutions to access central bank balance sheets.

This is precisely the opposite of the partial reserve banking model.

In a partial reserve system, the larger the scale, the more systemic risk is concentrated; here, the larger the currency scale and the higher the systemic importance, the closer its reserves are to central bank currency itself.

The endpoint is a currency that does not require central bank digital currency but possesses central bank-level security: issued by private institutions, programmable, internet-native, while having bankruptcy isolation and fiduciary licenses, and ultimately backed by the safest assets.

So what will happen to monetary policy?

When the partial reserve multiplier is no longer the primary transmission channel, price leverage still exists intact, and central banks remain responsible for setting policy rates.

Since stablecoin reserves are primarily composed of short-duration and overnight instruments, interest rates will be transmitted immediately and completely to the reserve base of the money supply, and the transmission may even be more direct than traditional bank credit channels.

What is fading is not the power of central banks, but the transmission mechanism of the money multiplier.

In two respects, the capacity of central banks may even be strengthened.

First, the transmission of policy rates becomes more direct; second, on-chain credit is transparent and observable in real-time, allowing central banks to see credit conditions directly without relying on lagging aggregate reports for inference.

The role of central banks will partially shift from operating the money multiplier to supervising a transparent machine credit market. This is a genuine expansion of regulatory functions, not a weakening.

The last distinction is crucial.

The reason base currency can be extremely secure is that it does not bear credit risk, and holding it does not generate interest.

Reserve earnings belong to the issuer and flow through the issuer to the stablecoin network ecosystem, but holding the currency itself is not an interest-bearing position. This is an intentionally set firewall to protect the security of base currency units.

Once holders begin to pursue yields, they are no longer just holding currency but borrowing into the credit market to obtain credit returns at the cost of bearing credit risk. This is an independent, actively chosen behavior.

The safety of base currency and the returns obtained from investing that base currency in the credit market are two completely different matters. Confusing the two will dismantle the entire argument for safety.

IV. Credit Market: Machine Underwriting, Agent Working Capital, and Prudential Layers

The previous section ended at a firewall: base currency is extremely safe because it does not bear credit risk, and holding it does not generate interest; the moment one starts pursuing returns, they cross the firewall into the lending domain.

This section discusses the system that grows on the other side of the firewall.

Credit will not disappear in a world of full reserve; rather, it will be re-established in a form that covers a broader scope, has more accurate pricing, and more transparent risk exposure.

The starting point is an observation that reframes the problem: long-tail borrowers such as small merchants, gig workers, and families, as well as future agents, have long been underserved by adequate credit services, not necessarily because their risks are too high, but because the cost of assessing each small exposure is often higher than the expected return of the loan itself.

Credit is restricted not often due to the quality of borrowers, but because of excessively high underwriting costs.

Once the assessment cost is driven toward zero, a large number of creditworthy yet long "unbankable" borrowers will finally become serviceable.

The force that drives down this cost is the data flywheel.

On-chain payment activities inherently possess structured, verifiable, and real-time characteristics, which are far richer than the lagging, fragmented records relied upon by traditional underwriting.

On-chain credit pools can also leverage oracles to bring off-chain facts into the system, including verification data from individuals, families, and businesses, as well as information from existing financial data tracks, credit histories, ledgers, and treasury system interfaces.

As treasury platforms, fintech companies, new banks, and enterprises migrate cash into on-chain currencies, the data will continue to thicken.

Since the network is global, the data will extend globally with every type of on-chain currency. This data will be input into real-time models with agent underwriting logic, forming a continuously compounding cycle:

Better data produces better models, better models produce better underwriting, and better underwriting attracts more activities and data.

Long-tail markets such as search advertising, content publishing, e-commerce, and software distribution are shaped by this recursive engine. Now, the same engine is starting to target credit.

Ultimately, what will be formed is a real-time, global, entity-authorized credit information system. In contrast, today’s credit agencies appear lagging, limited to a single country, and prone to errors.

A common objection arises here: does moving credit activities in the economy onto the chain mean exposing everyone’s financial lives to a public ledger?

The answer is clear: being on-chain does not equate to being public.

Selective disclosure and confidential computing technologies can allow contract states and positions to be encrypted and kept private by default, disclosed only through configurable, cryptographically enforced access policies, while protocol rules can still be deterministically executed on encrypted data.

An entity can prove key attributes to lenders, such as credit status, balance, and identity credibility, which can also be proven by oracles without exposing the original positions to competitors or the public.

The data flywheel can be both deep and private, with its confidentiality potentially stronger than today’s systems where intermediaries hold all information.

Regulators can see the authorized information, while competitors and the public see nothing.

As assessment costs decrease and data becomes abundant, underwriters themselves will also become agents, changing the economic structure of the credit market.

Agent underwriters will not tire, continuously optimizing towards the efficiency frontier, vying to underwrite exposures that past market structures could not cover, and relying on the continuously compounding data flywheel to enhance their capabilities.

The expansion of serviceable opportunities, compounded data advantages, and ongoing automatic optimization will combine to compress the marginal cost of borrowing, drive lending scale growth, and lower underwriting profit margins, just as machines compress stock trading spreads.

This will contradict a common intuition.

Typically, "cheaper, more abundant credit" is seen as synonymous with "more dangerous," a conditioned reflex left by the 2008 financial crisis.

But this system may become safer and more accessible even as credit becomes cheaper and more abundant, because the new efficiencies come from better information and better underwriting, not higher leverage.

This is the manifestation of "speed replacing leverage" on the credit side: scale growth relies on better underwriting and faster turnover, not on creating risky synthetic dollars.

What is Agent Working Capital?

The core idea is simple: agents can borrow money to fund the work they undertake, and the work they have already accepted can become an asset for lender financing.

This can be termed agent working capital, and the resulting asset can be called machine receivables.

It differs from traditional credit in that when a bank lends to a person, the biggest unknown is often whether the borrower is willing to repay, which is a question of human behavior.

Machine credit can partially eliminate this layer of uncertainty.

For example, if an agent has received a translation contract worth $10, it needs to borrow $4 to purchase additional computing power to complete the work.

The lender does not need to guess whether this agent "wants" to repay; they only need to price three specific things: whether the work will be accepted, whether the oracle will report truthfully, and whether the transaction will generate disputes.

Open-ended credit assessment thus becomes a shorter-term, clearly bounded question: can this work actually be completed?

One premise must run through it all: a single loan can be close to certain, but it will never be zero risk.

When a large number of similar loans are bundled together, systemic risks may still arise due to correlation.

The real change is not that risk has disappeared, but that risk can be observed in real-time and insured before problems arise, rather than rebuilding the scene after a system collapse.

The Logic of Collateral is Reversed

Traditional human collateral is usually an unrelated asset that courts take a long time to seize. Machine collateral, on the other hand, is exactly the opposite.

The first layer of protection for a loan comes from the reward generated by the work itself. The transfer of the reward can be completed on-chain, and when the work is settled, the lender automatically obtains priority repayment rights, with the recovery process completed through software rather than litigation.

More layers of protection can be set beneath this: collateral that agents pledge and can be forfeited, additional collateral, reputation tied to the creator, and the real responsible person standing behind the agent.

Once a problem arises, recovery will proceed in order: the escrow reward is first netted, followed by the forfeiture of pledged assets, then the shared insurance pool absorbs the tail losses, and finally, the remaining liability falls on the real responsible person.

The first three steps can be completed automatically in seconds. The last step is effective because at the end of the entire accountability chain, there exists a real and verified person.

This logic only holds in the short term.

The longer the loan term, the weaker the certainty.

For a signed work contract, financing for a minute's worth of computing power is almost a mechanical transaction; providing a few days of working capital increases certain risks; financing for a capability that has not yet been verified for months will reintroduce all traditional unknown factors, ultimately turning into ordinary credit.

Therefore, machine credit will not replace human credit but will form a new, near-risk-free short-term pricing floor.

Borrowing by individuals and businesses will be priced above this floor. This spread precisely measures the parts that machines have not eliminated: execution uncertainty, information gaps, and behavioral defaults.

At the very top of the yield curve, financing that relies on the founder's vision and long-term judgment for repayment still belongs to human credit and will long remain with humans.

Where Does the Funding Come From?

Retail investors can borrow funds into these markets just as they do today through new banks and exchanges' "wealth management" entrances.

This may further evolve into "agent wealth management": users provide funds, and agents continuously manage returns, risks, and redemptions, with an experience similar to a demand deposit account equipped with machine portfolio managers.

Corporate finance departments can participate in the same market through on-chain cash management, while institutions can bundle related assets into credit funds, with agents becoming the main interactive interface for all this liquidity.

Two things must not be confused.

First, the two types of returns come from completely different sources.

The returns generated from currency safety reserves do not belong to lenders but exist at the issuer and network ecosystem level; credit returns are the rewards obtained from lending, requiring active choice and bearing real risks.

Second, it is necessary to distinguish where the value ultimately flows.

Leading compliant issuers have begun to push a large amount of reserve income towards the ecosystem through partners and usage-based incentives, rather than directly paying it as currency interest. This ratio may continue to increase.

The resulting network tokens can be designed as stakeholder tools, allowing value to flow to validators, developers, and users.

The result is that the credit supply side itself will also be popularized and globalized, with long-tail participants able to both consume and provide credit.

However, a market that settles at millisecond speeds and is underwritten by machines will accumulate hidden exposures and collapse faster than any traditional institution can react.

Every financial system will fail; the real question is how it fails: is it like today, where problems appear late and the process is opaque; or does failure happen less often while the entire risk formation process is clearly visible?

Transparency will change the answer.

Markets can observe in real-time how exposures accumulate: every loan, every pledge, every relationship, while not having to expose private ledgers to everyone.

Businesses can keep positions private from competitors, while authorized regulators can see the entire system in real-time.

From post-event speculation of risk to real-time observation of risk formation, this is the fundamental change of the entire system.

But seeing a fire does not equate to being able to extinguish it. It is also necessary to write braking mechanisms into protocols, allowing them to operate faster than any committee's decision-making speed.

Rules are set by humans, and execution is completed by machines.

The most important brake is not a simple on-off switch but a dynamically adjustable knob: when too much capital concentrates into the same model, the same oracle, or the same computing power supplier, the cost of continuing to concentrate funds will automatically rise.

Risks will be gradually priced higher rather than suddenly hitting a wall.

Insurance must also become a real layer of existence, rather than a patch applied after the fact.

A shared insurance pool can be continuously funded by small deductions from each loan, with underwriters set up on top, and reinsurance covering the tail risks.

The novelty lies in the fact that premiums can be priced based on real-time observed risks, rather than relying on outdated historical averages. The health status of the insurer itself can also be continuously verified.

In the last financial crisis, the downfall of insurance giants was not merely about their eventual bankruptcy, but rather their long-term opacity, insufficient capital, and the inability for anyone to see the issues in a timely manner.

In the new system, risks may manifest before defaults actually occur.

Since the currency itself adopts full reserve, the base currency does not require a traditional safety net: there is no leverage embedded that needs to be dismantled, nor is there a run on the base currency that needs to be prevented.

This constitutes a true break from the banking model.

However, credit above the currency does not automatically ensure safety. The funding pool may still face concentrated redemptions, and collateral may still be forced to be sold off.

Thus, the real question is how to provide liquidity to the credit market during periods of stress, rather than how to provide deposit insurance for the base currency.

Possible solutions should primarily come from the private sector, including over-collateralization, reserve pools, reinsurance, and liquidity commitments made in advance by large fund holders.

As for whether the most critical infrastructure requires some form of public safety net, that remains an open question.

But even if public intervention is ultimately needed, transparency will allow for rescues to be more rapid, smaller in scale, and more targeted than in the past, rather than implementing large-scale rescues under conditions of insufficient information.

This system will also give rise to a new set of roles: underwriters, insurers, oracle providers, and funding pool operators.

These roles could essentially be undertaken by agents, but each agent must ultimately trace back to a real, accountable person. This is the premise under which regulation can still hold in the age of machine speed.

The potential regulatory structure can be divided into two layers: licensing regulation for significant participants whose size can impact the entire system, and constraints on long-tail participants through industry standards and self-regulatory mechanisms.

The role of central banks will also shift from maintaining the old money multiplier to jointly supervising these transparent markets with capital market regulators.

Relevant rules are far from complete: where should regulatory boundaries be drawn, how should borderless markets accept oversight from national institutions, and how to simultaneously prevent regulatory capture and regulatory gaps, all remain unanswered questions.

For the first time in history, institutions responsible for financial stability can take action based on a real-time, verifiable panoramic view of the system, allowing for gradual intervention rather than relying solely on blunt instruments.

This has a more solid foundation than today’s system.

Thus, the next question naturally arises: in a system that is so transparent, so globalized, and does not inherently belong to any nation, where does it actually "exist"?

V. Born Global

Three-layer Stack

The agent economy has a specific three-layer architecture.

The bottom layer is currency: software-based currency existing in the form of stablecoins, serving as a unit of account and final settlement medium.

The middle layer is the economic operating system: responsible for coordination, contracts, and value exchange, implemented through blockchain and programmable smart contracts with deterministic settlement finality.

The top layer is agent execution: the place where work is truly completed, driven by cloud software powered by AI foundational models and assistive models.

The most important aspect of these three layers is not just what they do, but where they exist.

Each layer is software, and each layer runs on the internet. The importance of each layer also comes from what it replaces.

Software-based currency replaces the national banking system that has mediated economic life for centuries. The traditional banking system is essentially bounded by nation-states, and cross-border transactions can only be stitched together through slow and expensive correspondent banking networks.

Software currency has no such geographical boundaries. No matter where it is held, it is the same money, and there is no need to inquire about the country of the counterparty at settlement.

The economic operating system replaces the legal and contract enforcement systems of nations. Coordination and trust have traditionally been constrained by judicial jurisdictions, as the courts and registration agencies that give real meaning to contracts are subordinate to sovereign states. The programmable settlement layer migrates some functions into deterministic code, ensuring that the rules are executed in the same way regardless of where the trading parties are located.

Trust comes from agreements, not from specific judicial jurisdictions. The agent execution replaces the most geographically rooted aspects: local labor and the companies that organize that labor. Execution completed by cloud-based AI models has no hometown. Anyone can invoke it from any location and at any scale. It responds to demand, not geographical location.

This is the core insight: each layer does not possess inherent geographical attributes because each layer exists in internet software, not within any national institution.

The economy assembled from these three layers also naturally inherits the borderless property of the internet. This can be termed "global by construction": globalization is not an added feature but a structural property determined by the constituent materials.

In recorded economic history, economic entities have always existed primarily within nations, and cross-border activities could only be retrofitted afterward. Now, for the first time, economic entities exist primarily within a global network, and it is the national framework that needs to be retrofitted.

No Single Native Jurisdiction

Every economic entity in history has been situated somewhere. Cross-border regulation has also been built on a rarely tested premise: that an economic activity occurs at a specific location, completed by a subject registered at a specific location.

Economic activities thus have a legal location, and jurisdictional issues unfold from there.

The agent economy may be the first economic system whose economic entities do not possess a native legal location. The reason lies in the work itself: work is executed by software agents, whose creators may be dispersed across a dozen jurisdictions, using models trained in one place, hosted in another, and called upon by counterparties in a third.

When such agents negotiate or settle, the classic question of "where does this happen" may have no clear answer at all. It is easy to label this situation as a jurisdictional vacuum, but the reality is quite the opposite.

Actions without a fixed legal location do not escape the law; rather, they may simultaneously fall under too many laws. Modern conflict of laws no longer requires a physical location. Mandatory rules attach based on the scope of impact and the location of the protected party.

An agent assembled from contributors across a dozen countries may simultaneously be subject to consumer protection laws where the client resides, data laws where the data subject is located, and tax rules where the market is situated, and these rules may conflict with each other.

The structural problem is not a lack of legal reference but rather the collision of too many jurisdictional claims without a clear legal location to help break the tie.

Meanwhile, a better coordinate system is emerging. Each agent is bound to a chain of accountability: it acts through credentials and wallets, tracing back through identity and trust structures to a verified, reputable real-world creator.

Compared to the world it replaces, this structure is actually clearer. In traditional correspondent banking, nominee structures, and offshore vehicles, the ultimate beneficiaries are often the hardest to see.

Thus, regulatory questions shift from territory to subject: who is the accountable entity behind this agent, and what obligations does it bear?

However, entity-based regulation does not mean that unified standards can be discarded.

If accountability is only attached to entities and not to locations, entities will actively choose verification locations. "Verified and reputable in a certain place" may degrade into "verified in a place that never raises difficult questions."

True constraints must come from the demand side.

The jurisdiction where users actually reside can require operators to meet recognized substantive bottom lines as a condition for market entry. Data regulation has already influenced global enterprises in similar ways, and recent tax agreements are also addressing digital value without fixed locations in a similar manner.

But there are two limitations.

First, being able to attribute actions does not equate to being able to enforce them.

Tracing actions back to a verified entity may yield a name but does not automatically provide remedies. Entities registered in jurisdictions beyond the reach of law enforcement, even with complete identity records, may refuse to assume responsibility.

The change is that when value flows on programmable tracks, enforcement can attach to the infrastructure layer rather than relying entirely on courts. Credentials can be revoked, balances can be frozen, and compliance status can become a condition for market access.

This leverage is faster than litigation, but it is also a double-edged sword, as it concentrates enforcement power in the hands of infrastructure operators.

Second, the identity layer that makes economic activities accountable may, from another perspective, also become a control mechanism.

The accountability engine and the scrutiny engine may be the same machine. Therefore, the identity layer cannot be established by a single operator creating a global unified registry but must possess diversity and portability. There needs to be competing credential issuers, credentials held by users themselves, and selective disclosure mechanisms that can prove "verified" without fully exposing identity to every counterparty.

Credential revocation must also be subject to due process constraints. Only with such design can traceability serve accountability without degrading into surveillance. This is an architectural choice that must be made consciously.

Compliance on the Margins

The systems in place worldwide to combat illicit funds are built on a quiet assumption: that money flows slowly enough, and that there are enough checkpoints along the way, allowing for post-hoc checks.

A cross-border wire transfer passes through a series of correspondent banks, each of which can only see a segment of the transaction; suspicious activity reports may only be submitted days later.

The old architecture is inherently opaque, fragmented, and oriented towards the past. Its maintainers often mistake this opacity for safety.

The agent economy can provide stronger control at the most critical nodes: the entry of subjects, the entry of funds into the system, and the moments when funds cross the boundaries of the regulated world.

These are precisely the places where the old system is most prone to failure. The old system lacks visibility and cannot intervene in a timely manner, while the transparent, identity-anchored, programmable settlement layer can partially remedy these issues.

Screening can be directly built into the settlement track, rather than being attached to every intermediary; it can become a gate that operates before settlement, rather than a report submitted after clearing.

Agents act through credentials traceable to verified entities, so a transaction can usually find a clear responsible party. This is difficult to achieve in the maze of traditional correspondent banks.

These capabilities do not require everyone’s financial lives to be published on a public ledger.

The solution remains selective disclosure: data is private by default, disclosed only with consent, and enforced through cryptographically secure authorization rules that grant reading permissions.

Regulatory and enforcement agencies can obtain certified visibility within the scope of authorization, while competitors and the public see nothing.

But a distinction must be made clear.

Programmability can ensure that rules operate at machine speed, but it cannot guarantee that judgments themselves are equally swift.

Real-time execution of established rules is the true new capability; real-time identification of new types of illicit activities remains a challenging adversarial problem.

The speed of machines may complicate this issue further, as the pace at which agents probe the boundaries of rules may outstrip human efforts to bridge the "gap between rules and intentions."

As a result, the problem of money laundering will not be fully resolved. Preventive measures at the boundaries will become faster and cheaper, while detection within the system remains a continuous arms race. New tools will be more powerful than today, but they will not bring about a final victory.

The key here is an architectural issue, not merely a monitoring issue.

A legitimate financial system must reserve space for true economic freedom and strong privacy: including self-custody, non-custodial wallets, and transfers that no operator can arbitrarily view or block.

This is not a loophole that needs to be closed, but a legitimate condition of a free society, the digital counterpart of cash.

Any standard that attempts to completely eliminate this space will ultimately construct a comprehensive control tool.

The correct approach is to set policies at the margins, at the boundaries where value and identity enter or exit the regulated world. Funding channels should be regulated, not the wallets themselves.

Illegal value is only truly useful when converted into real purchasing power, and this conversion often requires re-entering the transparent world. The redemption process of stablecoin issuers is an observable checkpoint that cash has never possessed.

The foundational layer must retain the freedom of use, while various jurisdictions establish their own control systems at the margins. If centralized control is directly written into the foundational layer, it creates a set of switches that can be captured, coerced, and abused.

A neutral foundational layer, by design, should not become a target for sanctions.

Recent legal practices have begun to acknowledge this. Regulatory agencies have previously attempted to sanction uncontrolled protocol code rather than sanction specific users, and such practices have ultimately been corrected.

The hardest capabilities to defend are often the most powerful. The same levers that can recover stolen funds can also cause wrongful seizures, large-scale automated misfires, and state scrutiny against legitimate but unpopular entities.

If a freeze order can be executed instantly and globally by the issuer under government pressure, without court intervention, its harm to freedom may exceed that of traditional bank freezes.

Freezing and recovery only have legitimacy under true due process, including cryptographic traces, automatic expiration without court renewal, multi-party authorization, and real and effective rights to appeal.

Thus, this architecture forces society to make a value choice, and this choice must be made publicly.

Retaining a bounded private inner domain means that some illegal value will remain beyond the direct reach of the state, just like in the cash era.

What this architecture provides is not panoramic visibility, but proportionality. The state can gain stronger tools than today: observable boundaries, screened funding channels, and marginal enforcement constrained by due process. The corresponding cost is the abandonment of the illusion of panoramic monitoring of the entire financial inner domain.

Multi-Currency and Invisible Foreign Exchange

Foreign exchange is the friction-filled seam between different national currencies. The slow and expensive nature of cross-border value transfer largely stems from this seam. A cross-currency payment requires a series of intermediary banks, each of which must pre-hold the currency positions of the other, each taking a profit cut, and each adding a day of processing time.

The way the agent economy dissolves this seam is by dissolving the premise behind it: each currency exists in its own national pipeline, and cross-border transactions must rely on intermediaries that span two systems.

As major currencies gradually go on-chain in the form of compliant, fully-reserved stablecoins, the relevant legal foundations are also being gradually established in different markets, and currency will gradually become an abstract layer.

An entity or its representing agent holds its national currency, the counterparty receives its own national currency, and the exchange process is settled atomically at the base layer, completing the settlement in one go and executing at the best price available in the market at that time.

Developers, agents, and end-users no longer need to consider the exchange process, just as applications on the internet do not need to consider every data packet and routing node when sending data.

The market for completing exchanges will not have a single mechanism but will form a diverse structure: price inquiry order books, automated liquidity pools, and other mechanisms competing and being routed to optimal execution.

This is a competitive market microstructure, not a monopolistic public utility.

The progress brought by the underlying settlement attributes is real. When exchanges are settled atomically, both legs of the transaction either settle simultaneously or do not settle at all. The time difference risk in cross-currency transactions, where one side's funds have already been transferred while the other side's funds have not yet arrived, can therefore be eliminated.

However, this does not eliminate all settlement risks; it merely shifts the risk elsewhere.

Fiat currency boundaries remain independent, non-atomic events, such as purchasing stablecoins with local bank currency or redeeming stablecoins for bank currency.

The reliability of non-USD stablecoins still depends on reserve depth and redemption liquidity.

What the agent foreign exchange layer truly accomplishes is the transformation of opaque counterparty risks in the intermediary banking system into transparent, priceable anchoring and redemption risks.

This is a visible, measurable risk, rather than a risk hidden in a series of foreign bank balance sheets.

The more significant commitment lies in reaching the long tail of currencies. Currencies from countries like Paraguay, Kenya, or the Philippines currently have high transaction costs globally, not necessarily due to a lack of demand, but because the fixed operating costs of the intermediary banking system exceed the value of the relevant transaction flows themselves.

Once currencies go on-chain, these fixed costs may drop to near zero, making currencies that were previously unserviceable now serviceable.

This is entirely analogous to how the internet compressed distribution costs to near zero, allowing niche products to emerge in the long tail.

However, there is an important limitation: liquidity is not content. Continuously making markets for a niche currency requires real capital inventory and entails actual risk. This variable cost will not drop to zero. Therefore, the currency long tail will extend significantly but will not immediately become complete.

In practice, the vast majority of exchanges will still use USD as a bridge. When one small currency is exchanged for another, it is more likely to first settle as the local currency against USD and then settle as USD against another local currency, rather than establishing a native trading pair directly. The reason is that concentrating liquidity in a single carrier currency is far more efficient than maintaining a fully networked market that requires astronomical numbers of trading pairs.

This system is multi-currency at the endpoints, while the intermediate pipelines may be highly dollarized. The user experience is genuinely local currency to local currency, but the value will pass through a USD hub in the middle.

This hub will not disappear, and its continued existence will also impact currency sovereignty. The same architecture will also reshape the treasury management of enterprises and financial institutions.

The corporate treasury existing in on-chain currencies will no longer be a collection of accounts scattered across different countries and banks that require effort to aggregate, but will become a single, global, round-the-clock operation managed by policy rules.

Idle funds may become a historical relic. Balances will continuously be swept into yield strategies or credit markets, rather than waiting for nightly batch processing. Humans are responsible for setting programmable guardrails, while agents execute within those guardrails.

Of course, the friction brought by the last mile and foreign exchange depth still exists. Global treasury entering or leaving the local currency system still requires passing through traditional edge licenses, banking partnerships, and liquidity.

Reshaping Currency Sovereignty

When currencies begin to flow on a neutral global software layer, one of the first concerns that arises is currency sovereignty.

If value can be settled anywhere in seconds, and a certain currency dominates cross-border routing, it seems that different countries, especially small ones, are relinquishing control over their currency affairs.

However, this intuition confuses two things that are being architecturally separated: the track of currency flow and the currency flowing on that track.

The protocol layer is designed to be governed by multiple stakeholders, independent of jurisdictions, and not belonging to any single country. However, the currency flowing on the protocol is still anchored to specific jurisdictions.

A compliant, fully-reserved stablecoin is still a debt instrument denominated in a sovereign currency of a certain country and issued according to that country's laws. The protocol can remain neutral, but the currency does not lose its jurisdictional affiliation.

This distinction is crucial.

Protocol neutrality does not equate to settlement asset neutrality.

USD stablecoins are regulated liabilities issued by entities responsible in specific jurisdictions. They effectively carry a shutdown switch of a foreign government, which may freeze or isolate them by a government that is not the holder's home country.

However, protocol neutrality is precisely what makes it more valuable.

Only if the foundational protocol remains neutral can other countries potentially issue their currencies onto the same track, thereby reducing dependence on assets with foreign switches.

The neutrality of the foundational layer is the prerequisite for the re-emergence of currency sovereignty.

Thus, bringing one's national currency onto the chain in the form of compliant, fully-reserved instruments is not necessarily a concession of sovereignty; rather, it may be an upgrade of sovereignty.

A currency that could only reach the global stage through a slow, expensive network of intermediaries can become globally programmable and directly usable by anyone.

The core lever that constitutes currency sovereignty—the power to set the price of funds in local currency—remains in the hands of central banks.

What changes below interest rates is the transmission pipeline, not the pricing power.

This will also redefine sovereignty itself.

The old definition of sovereignty is territorial: controlling tracks, guarding borders.

A more enduring definition of sovereignty may be competitive: operating a sound currency that can win users based on credit. Different currencies compete through credibility rather than relying on borders to enforce loyalty.

This upgrade is genuinely feasible for currencies with deep markets and credible institutions, but it is challenging for the smallest and most fragile currencies.

These currencies struggle to attract issuers and liquidity because few are willing to hold them. However, the on-chain track at least lowers the threshold for small currencies to achieve basic usability, an opportunity that the intermediary banking era never provided.

The gap between sovereign currencies existed long before these tracks appeared. The real question is whether the new tracks will continue to amplify the gap or provide a first affordable path for weaker currencies to catch up.

The most challenging issue is digital dollarization.

If residents of weak currency economies hold USD stablecoins, making it as simple as using a chat application, currency substitution will become unprecedentedly frictionless. This could siphon demand away from the national currency, disintermediate local banks, and weaken a central bank that can issue local currency but cannot issue USD in residents' hands.

This architecture will impose more direct discipline on policy: when a sound currency becomes a frictionless option, unsound policies will immediately incur clear and visible costs.

However, this discipline is asymmetric. It will only fall on weak currencies and not equally on the issuing countries of dominant currencies, as the latter's currency is itself the asset everyone escapes to. This excessive privilege is real and predates on-chain currencies by several generations. It stems from network effects rather than the protocol itself.

Digital dollarization must be managed through a "policy at the margins" approach, rather than expecting it to disappear on its own. Capital flow measures, holding limits, and exchange rules can be employed, and these rules can be expressed in code and enforced at the funding entry and exit channels. These tools remain localized. The less permissioned the internal space, the more likely marginal controls will have vulnerabilities.

However, compared to traditional execution models, they are at least more precise and easier to observe. The real threat is ceding the territory of one's own currency directly to foreign currencies. This is precisely the important reason for issuing domestic stablecoins as early as possible.

A world composed of sound on-chain currencies, deep liquidity, and effective boundary tools may be more stable and even improve overall welfare.

However, the history of capital account liberalization has shown that the real risk of economic rupture often occurs during the transition phase. When the speed of currency substitution exceeds the pace of institutional and buffering mechanism construction, risks can concentrate and explode.

Moreover, these trajectories are coming, regardless of whether any sovereign nation chooses to act.

Therefore, the real choice is not "whether to transition," but rather "to undergo a managed, sovereign-led transition" or "to be forced to accept an unmanaged transition." Sovereignty in the agent economy is being reshaped, and only by managing currency to a level worth holding can sovereignty truly be preserved.

Interoperability and Migration from Traditional Systems

The agent economy will not emerge from a vacuum. It will be built on a planetary-level existing payment infrastructure, including bank transfers, card organizations, wire transfers, and electronic currencies. This system has been operating for decades and will not be completely eliminated in the short term.

Thus, this will not be a replacement completed overnight. The crypto industry has repeatedly told this story and has repeatedly misjudged it.

The more realistic process is that the new settlement base gradually takes shape beneath the new value created in the agent economy, while the old tracks continue to carry existing value, with both connected by bridges, slowly moving boundaries in one direction.

On-chain systems will become the core.

They will become the native place for the generation and settlement of new value, especially carrying those new values or value flows that the old tracks were never designed for: cross-border value, programmable value, value flowing continuously around the clock, and machine-to-machine value exchanged at scales and granularity that human-paced tracks cannot accommodate.

The traditional system will gradually become marginal, providing the last mile for on-chain value to reach terminals that have not yet migrated.

Of course, in many domestic low-friction scenarios, the traditional system has not stagnated. Real-time bank transfer systems can already provide instant, cheap, and final domestic settlements.

The issue is not whether these tracks will disappear, but rather that the frontier of new value formation will increasingly exist natively on-chain, while the traditional world will continue to carry existing value for many years to come.

The two worlds are connected by bridges, but the nature of these bridges must be accurately described. Bridges remain intermediaries, the very type of trusted parties that on-chain models attempt to minimize. What is truly happening is not the complete elimination of trust, but rather the transfer and compression of trust.

Here, bridges should not be those fragile cross-chain token bridges that have caused significant losses in the past, but should be closer to regulated clearing facilities: licensed, capitalized, and possessing traditional liability systems and disposal arrangements.

Bridges are systemic nodes, and their governance and disposal must be designed with the seriousness of a clearinghouse.

The most challenging part remains the last mile, as converting on-chain balances back to local currency has long been the weakest link in the crypto industry. Stablecoin balances are indeed more flexible than pre-deposited accounts scattered across agent banks, but the terminal remains a difficult task.

In a country, cashing out funds to local accounts still relies on banking partners, licenses, and the depth of local liquidity at the time of exchange. These conditions are often most scarce in emerging market corridors where cross-border demand is strongest.

On-chain itself will not eliminate the agent bank problem at the terminal; it will only concentrate the problem at the last mile. A unified network truly changes is that the fixed costs of licenses and partnerships can be diluted. A single network can distribute partnerships, licenses, and local liquidity across multiple jurisdictions, without requiring each company to establish bilateral relationships in every country.

Consumer and merchant interfaces will also connect to the on-chain core through an "embrace and extend" approach. Tokenized card vouchers can allow on-chain funds to reach existing merchant acceptance networks, but related transactions still operate on card organization tracks, and the rules and fee structures do not change as a result.

Therefore, it is necessary to distinguish between the migration of the settlement layer and the migration of the acceptance layer. The former is likely to occur, while the latter faces greater controversy and may never fully happen.

A lasting equilibrium may be that stablecoins only provide funding for the existing card networks, while card organizations continue to control the acceptance interface and most of the rents.

This is a real possibility. However, as the agent economy unfolds, more and more services will directly use on-chain currency for fulfillment, and the acceptance layer tracks will ultimately yield to the on-chain environment.

The key fact that links the entire migration is that the agent economy creates net new demand. This is the value exchanged between software agents, with no existing tracks specifically designed for it.

It rides on the long-term growth of AI and does not need to rely on zero-sum replacement of traditional institutions. On-chain systems do not need to win in competition with card organizations first to become the foundation of the agent economy. They only need to be the place where the agent economy can operate natively.

Coexistence will last a long time, but the new base is being laid, and the boundaries between the two worlds continue to move in the same direction.

The Dual Nature of Equality and Its Counterforce

Every attribute that makes the agent economy powerful is a double-edged sword.

The lack of borders allows creators in small markets to reach the whole world, but it also lets competitors from around the world enter that small market.

Permissionless removes the gatekeepers blocking unbanked individuals, but it also dismantles the gatekeepers protecting local established businesses.

Near-zero marginal costs enable anyone to serve everyone, but they also allow the most capital-rich participants to be the first to serve everyone.

This design inherently possesses duality: the same characteristics can generate strong equality effects while also driving high concentration.

The equality aspect is real, and in some places, it is even absolute.

Previous rounds of digital opportunity rhetorically promised equality for all, but in practice, there were barriers everywhere: distribution required approval from app stores, payments required bank accounts, entry into capital markets required qualified investor status, and the global financial system operated on a network of agent banks that excluded large populations.

The on-chain agent economy can eliminate these gatekeepers at the protocol level.

Holding sound currency, participating in transactions, providing or using credit, and selling work produced by agents to the global market may no longer require prior permission.

For those truly excluded, this represents a difference from zero to one.

A means of storing value independent of collapsing domestic currency is especially important for populations in high-inflation economies who have long been underserved by financial services.

But the counterforce is equally real.

Eliminating friction caused by borders, currencies, and local licenses is a two-way street, not just a one-way gift to marginal areas.

The world entering small markets is as easy as small markets entering the world. When global competitors arrive, they often have more capital, stronger agents, and lower costs.

Local participants previously relied on distance, language, currency, and regulation to form protective survival. Once this layer of protection is stripped away, they will face a winner-takes-all competition in which they are naturally at a disadvantage.

Moreover, winning takes all here is not a marginal risk but a default baseline.

Currency is the most typical network effect product; a dominant settlement asset will naturally become the default choice; cutting-edge AI capabilities are also subject to extremely high capital thresholds, with scale requirements far exceeding those of the open Web era.

The open Web ultimately left only a few giants, and the same gravitational pull will also act on the agent economy.

The weights of the two outcomes are not equal.

Concentration is the default result, naturally produced by these technical characteristics and reinforced by the deeper laws of platform economics.

Equality, on the other hand, is an alternative that needs to be actively constructed.

It will only occur when certain specific conditions are consciously created: building open infrastructure in the core layer that could originally charge tolls; designing ownership distribution rather than allowing capital gravity to dominate; using policies to prevent gatekeepers from re-forming at new barriers, such as model layers, dominant issuers, identity layers, and bridge layers.

Equal access does not equal equal outcomes.

Creators from anywhere can enter the global market, but that does not mean profits will ultimately flow to anyone. A more honest statement is that the economic baseline may be raised, but the gap may not necessarily narrow.

The final outcome is a choice made by ownership structures and policies together. Technology has for the first time brought better outcomes into the realm of deliberate design.

An economy that is born global cannot be merely a technical or economic fact; it is inevitably also a geopolitical fact.

From this, three inferences can be drawn.

First, infrastructure must remain technologically neutral and be governed by a broad range of stakeholders, rather than belonging to any single country. A global economic operating system controlled by a single power is difficult to gain trust and maintain stability.

Second, when a major power establishes its currency as a regulated digital currency, the consequences will be profound and lasting. It will accelerate the migration of value to these tracks and force other governments to choose to compete, adapt, or resist.

Third, the increase in the speed of economic flow will also raise the cost of conflict. A world where commercial relationships intertwine at machine speed will embed reasons for participants to be unwilling to tear apart the system within economic pipelines.

This is the most hopeful of the three inferences, and also the most uncertain.

VI. Supply Side: From Subscription to Pay-Per-Use

The agent economy requires a supply side: a universe of services that agents can call upon, hire, and pay for.

This supply side will form in two waves.

The first wave is to package existing things: software, data, and network services expose themselves to agents through command line interfaces, machine-readable packaging, and skill layers, and are repackaged, measured, and priced for machine consumers.

The second wave is to create things that did not exist before: specialized agents that delve into a certain field and sell their work directly to the market.

The most profound economic rupture is very simple: the unit of value shifts from "access rights" to "work." Just this change alone is enough to redefine the pricing model of the entire software industry.

For the past thirty years, the mainstream business model in the software industry has been subscription-based by seat: users pay ongoing fees for access rights for one person sitting in front of the tool. The agent economy will gradually dissolve the seat model.

Consumers will no longer be the ones occupying licenses but the agents executing tasks; what is purchased will no longer just be access rights to tools but the work itself.

What truly disappears is the seat as the core pricing axis, not the subscription model itself. The billing unit will shift from "how many people can log in" to "how much work has been completed." Around this new unit, various forms of business structures will simultaneously take shape.

Pure pay-per-use billing is suitable for sporadic and exploratory work; subscriptions that include usage quotas can incorporate a certain number of work units within fixed commitment spending, with excess billed separately, thus restoring budget predictability needed for finance and procurement; outcome-based billing is suitable for scenarios where results can be clearly defined and measured.

These are not mutually exclusive predictions but a form of pluralistic equilibrium.

The same logic will further enter between the agents doing the work and the foundational models driving them.

True value migration is happening here. As dedicated AI agents proliferate, buyers are increasingly purchasing outcomes from agents rather than directly buying tokens from models. Agents absorb the intelligent costs as sales costs and arbitrage between different foundational models, completing tasks at the lowest quality-allowed cost.

This change is no longer just theoretical. Model routers—assigning each request to the most suitable system based on cost, latency, and quality—have transformed from optional tools to critical infrastructure within a year. Many enterprises are now running multiple models simultaneously in production environments. The routing systems are said to significantly reduce costs while maintaining quality.

The drive for routing is extremely strong due to the vast price differences between different models: the cheapest production-grade models may cost only a few cents per million tokens, while the most advanced frontier models can be orders of magnitude more expensive.

Assigning simple tasks to expensive models is purely wasteful. A whole set of cost governance disciplines is forming around these issues, along with a corresponding new vocabulary. "Tokenmaxxing" describes a failure mode: systems optimizing to consume tokens rather than to create value.

Corresponding control measures are also emerging rapidly. Large enterprises with rapidly rising internal AI billing have begun to enforce relevant measures. The unit economics of agent work is not a distant abstract concept but is being rapidly shaped.

The ultimate conclusion can be distilled into one sentence: models become cost items, and agents become businesses.

Value is more likely to accumulate at the layer that holds customer relationships, proprietary contexts, workflows, and outcome responsibilities, rather than accumulating in the raw intelligence itself. This aligns with the pattern of every previous round of platform transformation: above commoditized inputs, a new layer capturing value always emerges.

However, there is also a serious downside.

Owners of frontier models will not passively wait to be commoditized. In truly challenging capabilities at the frontier, if a certain model is clearly stronger, routing will degrade into "directly using the best model," and model owners will retain real pricing power as a result.

At the same time, they will also integrate upwards into the agent and application layers, holding both models and customer relationships. This could ultimately form a barbell structure. A large amount of commodifiable work in the middle will be arbitraged and profited by dedicated agents between interchangeable models; a small portion of frontier work will be retained by model owners as rent, directly competing at the agent layer.

The boundary between the two will continuously rise with model advancements. Yesterday's frontier becomes today's commodity, continuously entering the arbitrageable middle ground, while new capability frontiers will open higher up.

Beneath the pricing of work lies the settlement of work. A promised internet dream of thirty years may finally be realized here: micropayments.

Micropayments refer to charging for extremely small units of value, often less than a cent. The consumer internet has long promised that it would arrive but has never truly materialized. The common explanation is that settlement costs are too high, but that is only half the problem. A more fatal barrier is the psychological transaction cost for humans: people are reluctant to repeatedly judge whether something is worth a penny, and the cognitive friction of decision-making far exceeds the price itself.

The agent economy can simultaneously dismantle both types of barriers. Programmable rails can settle a few cents at low cost, while consumers are no longer people who hesitate over a penny but machines without such psychological resistance.

Thus, micropayments may finally be on the verge of becoming a reality.

The most suitable scenario for this is not the repeatedly failed content access scenarios of the past, but labor: measuring and exchanging small units of work between agents.

Friction has not completely disappeared but has shifted. The judgment of "is it worth it" will be recoded into budget strategies and work value assessments in software. Measuring millions of sub-cent events will also incur accounting and observability costs.

However, these costs can be diluted: a single strategy can govern millions of decisions; they can also be aggregated: a large number of micro-events can be netted, packaged, and settled collectively without needing to complete final settlements one by one.

Thus, the major historical barriers collapse and are replaced by a new, lower cost that can be diluted. This is what makes labor at a penny granularity economically viable for the first time.

When the consumers of services are machines rather than attention-limited humans, the services themselves will also begin to transform.

An increasing number of services, including consumer-grade web services, will be restructured with the default primary users being agents: callable by machines, priced per use, and stripping away interfaces, funnels, and advertisements designed to capture and monetize human attention.

The market that these services can reach will expand dramatically, as an agent can concurrently consume thousands of services around the clock, no longer limited by human attention bottlenecks.

This restructuring will initially be partial and gradual. Existing enterprises may first add a layer readable by agents outside the human interface, and only after a long time will they truly dismantle the original human interface.

Many companies will also resist, keeping external agents at bay or only allowing their own agents to enter, as today’s internet is a bilateral market worth hundreds of billions of dollars, and existing enterprises have ample motivation to protect it.

Here, a real revenue gap will also emerge: what the advertising internet sells is essentially human attention.

Machine consumers have no attention and will not watch advertisements, potentially siphoning off the revenue model that supports half of the internet. However, the gap and the patch are essentially the same thing.

When agents bypass attention-dependent monetization models, the natural successor will be directly measuring and charging for the services consumed by agents, which is pay-per-use and micropayments. Pay-per-use by agents is likely to become the next generation revenue model after the attention economy and advertising economy.

But all of this hinges on a layer often overlooked by optimistic narratives, and recent experiences have shown the cost of lacking it. If work can be precisely measured, and agents can hire other agents, tools, and models, then the surface of what can be spent is essentially boundless. The agents that truly incur spending are not the entities ultimately holding the bill.

Machines do not instinctively fear spending money. When they fall into a negative feedback loop or devise overly aggressive plans, they can generate massive bills in a very short time. Large enterprises have already seen internal AI usage exceed budgets and have begun responding with hard caps on single tools and centralized control.

Thus, the agent economy must have a governance layer to operate normally.

This governance layer needs to include spending caps and budget rules, limits passed down along the delegation chain, real-time measurement, anomaly detection, and manual approvals for significant actions. It can be understood as the agent version of cloud cost management and corporate card limit control.

It will gradually become an independent product category. This does not undermine the argument for the agent economy; rather, it complements it. A service that can be priced and called by machines must also expose strategies, identities, and authorizations simultaneously.

Next to the micropayment rails, there must be a budget and authorization rail, anchored to the same accountable identity that runs through the entire system.

Finally, it is necessary to distinguish between two easily confused time scales.

The speed of development of costs and capabilities is indeed rapid: token costs have dropped by several orders of magnitude, cost control technologies are maturing on a monthly basis, and the most advanced scenarios—first in software development and other digital-native work—are quickly entering the mainstream due to their higher credibility, easier integration, and more measurable outcomes.

However, the speed of enterprise adoption involving significant responsibilities and strict regulations is much slower and is fundamentally independent of the cost curve.

What hinders adoption is not unit economics, but the integration with recording systems, security and compliance reviews, the immaturity of governance layers, who bears responsibility when agents make mistakes, and the resistance of procurement systems to new pricing methods.

Both time scales will coexist. The direction changes quickly, enabling economics to arrive rapidly; but achieving widespread adoption in significant areas will still take years and will be unevenly scrutinized.

However, the overall direction is already very clear: value units are shifting from access rights to work, value is flowing to agents that hold outcomes, intelligence is gradually becoming commoditized for most tasks; micropayments are first realized in the labor market as buyers become machines; the internet begins to reshape around new agent customers; and governance layers emerge, making autonomous spending safe.

This will not be a clean and tidy switch but a real repricing of the way work is bought and sold in the economy.

VII. On-chain Companies

As AI takes on more and more work for companies, companies themselves also need a new space to exist.

As a company's labor is increasingly completed by software agents, and these agents can hold value, sign contracts, coordinate with each other, and take action globally, a foundational economic base is needed to enable these behaviors to truly occur:

Currency can be held and transferred by programs; decision-making and execution rules can be expressed in software; coordination between humans and agents can be recorded and executed; and the company's external economic relationships can be completed at machine speed.

This economic base is the on-chain economy.

Agent companies and on-chain companies are two sides of the same entity. The agent side describes who is doing the work, while the on-chain side describes in what form the work exists.

This is also the core of the entire discussion: the agent economy is the on-chain economy.

The two are not two adjacent trends that may intersect in the future but the same phenomenon. An economy run by software agents must operate on software currency, software contracts, and software governance; otherwise, it cannot truly function.

This does not mean that every company will dissolve into a collective governed by tokens. The future of companies will be a hybrid form, developing along two parallel paths.

The first path is the evolutionary path.

Existing companies, including ordinary Delaware C corporations, will gradually tokenize equity, map governance mechanisms to on-chain mechanisms, and migrate part of their operational entities to programmable infrastructure while retaining familiar legal forms.

This change has already begun to appear: the attitude of securities regulation is shifting, share registration and transfer agents are starting to pilot on-chain registration, companies connecting relevant infrastructure are emerging, and some corporate laws are beginning to allow distributed ledgers to directly serve as shareholder registries.

But this will be a long path, as it is constrained by the slowest and most conservative institutions in the financial industry.

The statutory laws governing ownership authority registration methods, the central securities depositories relied upon by public stock settlements, the yet-to-be-defined tax treatments and auditing standards, and the cautious attitudes of boards and legal advisors will all slow down the transformation.

This transition may take one to twenty years to unfold, rather than being completed within a single market cycle.

The second path is the native path.

Highly agentified new companies can adopt a native architecture from day one, making governance, treasury, and digital tokens the core primitives of company operation rather than additional features on top of traditional companies.

These two paths correspond precisely to the changes occurring on the software supply side: existing enterprises are evolving their current products into services that agents can consume, while new participants are building agent-native products from scratch.

Native builders without historical baggage will take the lead in demonstrating new models, pushing existing enterprises and regulators forward.

However, even native companies cannot escape the law simply because they were born in software. Legal personality, limited liability, the ability to enter into contracts, and the right to sue are all granted by sovereign law, not by a single on-chain deployment transaction.

A smart contract system that is not registered in any jurisdiction that recognizes its status may be legally deemed an unincorporated association or a general partnership.

This means that participants may need to assume unlimited personal liability. Existing token governance collectives have encountered this issue in real-world litigation. Therefore, truly operational native companies still need to register through a thinner legal shell, such as a specially designed DAO LLC or similar legal entity, bridging the code into existing corporate law containers.

The real change is not that companies can exist outside the law, but that the proportions of various parts within the company have reversed.

The legal shell becomes thinner and closer to a formal container; the operational entities—including treasury, payroll, contracting, and governance execution—exist more on-chain and become more substantial.

This is a genuine transformation in how companies operate, but the role of sovereignty in granting legal existence to companies has not disappeared.

Entities entering these new forms will also change. On-chain companies combine traditional governance structures—a registered legal container with limited liability—with tokenized ownership. Equity or digital tokens can carry contractual rights to income, voting rights, participation rights, and other utilities within the entity. Companies will also replace a large amount of non-digital infrastructure with programmable infrastructure: on-chain treasuries operating under clear policy constraints, continuous rather than periodic auditability, and default global interoperability.

But in this evolution, three things must be clearly distinguished.

First, tokens can today represent a share of equity, which is a reality that has already been achieved.

Second, whether blockchain can become an authoritative record, that is, a legally valid formal register of "who owns what," remains an evolving statutory issue. This step will only be completed when corporate law formally recognizes on-chain ledgers as statutory shareholder registers.

Third, whether property rights can achieve truly legally final on-chain settlement, rather than merely achieving technical finality that maps stock tokens, remains fundamentally in the vision stage.

Before blockchain becomes a statutory register, a tokenized stock actually exists on two ledgers simultaneously.

Those mechanisms that seem less glamorous but are truly difficult, including proxy voting, information disclosure, dividends, withholding taxes, lock-up periods, transfer restrictions between qualified investors and retail investors, and regulated custody, are precisely the slowest parts of progress.

"Being able to represent" has been achieved, "becoming an authoritative record" is the current frontier, and "achieving complete legal settlement finality" is the ultimate goal.

The three cannot be conflated. The most profound changes occur at the governance level.

As more and more decisions within companies are made jointly by humans and agents, companies need a tamper-proof record that both parties can read, write, and act upon with equal authority. Without such a record, the governance structure will crack at the seams between humans and machines.

A cryptographically provable authoritative ledger can become this shared source of fact: from the execution of agent actions, all the way to token and equity holders, and ultimately extending to the boundaries of the board.

It will become the backbone of human-machine collaborative governance, but the capabilities of the ledger must be accurately described.

It can permanently and verifiably confirm what happened, in what order it happened, and by whom it was completed. This brings enormous benefits in terms of evidence preservation and non-repudiation, potentially ending a whole class of factual disputes. However, it cannot determine whether an action was authorized, whether it was within the entity's authority, whether it was sufficiently prudent, or whether it met fiduciary duties.

A self-dealing or unauthorized transaction that can be perfectly proven may still constitute a violation. Permanence may even firmly fix a wrongful act in the record. The ledger is a better witness but not a better trustee.

The duty of care remains a human obligation. Machines cannot truly bear fiduciary duties, cannot be questioned about subjective intent, cannot be disqualified for disloyalty, and cannot bear personal legal liability. Therefore, when agents make or execute governance decisions, the obligation does not disappear but falls on the individuals who design, configure, authorize, and are responsible for supervising that agent.

Machine governance does not remove accountability from humans; rather, it clarifies accountability on the designers and supervisors, adding an obligation to supervise the agents on top of the existing duty of care.

"Humans are in the governance loop" thus takes on a more accurate meaning: humans bear obligations, the ledger keeps records, and agents act within the limited authority granted by humans.

The same precision applies to the claim that "contracts become programs."

Contracts are ultimately a set of rules for decision-making and execution. These rules will increasingly be written in code and automatically executed by machines between companies and agents, and between agents.

In this sense, contracts will indeed become programs.

But code executes literally and completely, while legal contracts often intentionally remain incomplete.

Legal contracts rely on courts to interpret intent beyond the literal, and depend on a series of legal principles that allow business to operate, including mistakes, fraud, coercion, impossibility of performance, and consumer protection, employment law, and competition law that parties cannot exclude through contracts.

Therefore, the true picture is a two-layer structure.

Programs are the fulfillment and settlement layer, responsible for automatically executing agreements along high-frequency, low-ambiguity paths. In terms of transaction volume, this will cover the vast majority of activities.

Legal contracts remain governance documents, responsible for handling interpretation, defenses, and mandatory laws, addressing the few paths where code and real intent diverge.

Automation applies to fulfillment, not meaning. Classic cases have already emerged. A well-known early smart contract system was attacked; the code executed exactly as written, but the community still believed the outcome contradicted the true intent and ultimately chose to roll back. Even the most code-native participants will place the final judgment outside the bytecode when the execution result conflicts with the actual intent.

Contracts will become programs in terms of fulfillment, but remain legal documents in terms of governance. This is why on-chain companies are best understood as: core integrity is ensured, while the margins are responsible for adjudication. The deterministic core handles a large number of low-ambiguity activities, including transfers, payments, ownership unlocks, and simple conditional logic.

Its degree of automation and auditability is unmatched by any traditional human backend. In these scenarios that occupy the majority, trust is indeed significantly minimized.

The margins handle the few disputed scenarios, including introducing off-chain facts through oracles, resolving disputes through arbitration, and intervening with human mechanisms when "code is correct, reality is wrong." This margin is re-intermediation, not complete de-intermediation.

Oracles are themselves a trusted party for the facts they report; the arbitration layer must be operated by certain entities; an intervention key capable of reversing core results is, by definition, a centralized node that can be captured, coerced, or compromised.

Whoever holds this key holds the company in extreme situations.

The real victory is not the complete elimination of intermediaries, but making intermediaries transparent and competitive. Intervention rights should be held collectively by multiple parties, constrained by time locks, and recorded on the ledger, rather than relying on a silent administrative key.

The rules and incentives of the arbitration layer should be publicly visible, and the sources of oracle information should also be verifiable. The core remains intact, while the margins mediate through transparent and accountable means; this mixed structure is the true architecture of future companies. It is more honest than the fantasy of complete trustlessness and more reliable than defending traditional opaque intermediaries.

Finally, two warnings need to be added.

On-chain governance will not automatically be more democratic than traditional equity registration and proxy voting. One token, one vote is structurally oligarchic, concentrating power in the hands of large holders and insiders; participation rates often remain low over the long term, while also creating new attack surfaces such as vote buying and governance arbitrage.

Secondly, the timeline should not be exaggerated. The direction is real, and the infrastructure is being rebuilt, but the transformation of existing companies remains a decade-long project, constrained by the most conservative institutions in the financial industry.

However, the shape of the endpoint is already quite clear. Companies will acquire a new material form: currency, decision-making, coordination, and external economic relations will all be expressed in software; the external layer of the company will be wrapped in a thinner legal shell granted by sovereignty; internal records will be on a provable ledger, which becomes the governance backbone at the human-machine seam; contracts will be fulfilled by programs and governed by law; core integrity and margin accountability will maintain a balance.

Some companies will reach this endpoint through evolution, while others have existed within it since their inception.

The two paths will progress in parallel but ultimately point to the same conclusion:

Agent companies and on-chain companies are the same company because the agent economy is the on-chain economy.

VIII. Concentration of Influence and Power

The agent economy places the greatest opportunities and the heaviest risks of this era in the same machine.

This is not two futures that can be freely chosen, but rather the results that the same system may simultaneously produce, with the final balance yet to be determined.

What truly needs diagnosing is: what mechanisms allow each outcome to prevail, and what forces are tilting the balance.

First is the labor issue.

Automation does not necessarily destroy jobs net. The so-called "fixed total labor" hypothesis has been repeatedly proven wrong by two hundred years of economic history. Old jobs are replaced, but new tasks continually emerge; comparative advantage also means that even if an agent is absolutely stronger in all things, it will still only have a relative advantage in certain tasks, and humans will still have work to do.

But comparative advantage only discusses relative productivity and does not address price issues. Humans can continue to be employed in the jobs where machines are relatively weakest, but the wages for these jobs may drop to levels that cannot sustain a family. This state may statistically still be considered "full employment," but socially it could be a disaster.

Therefore, the truly serious issue is not the quantity of employment, but the share of labor in total output and the market-clearing wages for human work.

This concern will hold under three conditions.

First, new tasks continue to emerge as they have historically, but the speed at which software occupies new tasks outpaces the speed at which humans can retrain and enter new fields. The historical lag between replacement and reabsorption has been compressed to near zero. Agents are not just more efficient workers in today's tasks; they may also be faster occupiers of tomorrow's new tasks.

Second, the marginal price of agent labor will continue to decline with reasoning costs. At an increasingly broad frontier of tasks, machines will become cheaper on the margin and will drag down the market-clearing wages for human labor along with it.

The third, and truly different from previous technological waves, is that capital can fund its own expansion. A loom does not generate income on its own, nor does it design the next generation of machines; however, intelligent agents can create surplus, finance more intelligent agents, and even participate in writing their successors. This can form a self-reinforcing cycle: capital creates software, software generates more capital, and capital further expands the scale of software. As a result, the share of capital may self-reinforce, no longer constrained by the traditional diminishing marginal returns.

When these three conditions are met simultaneously, the main channel through which most people have historically shared economic output—labor—may narrow to a trickle.

However, even if all three conditions are met, the outcome may not necessarily be widespread poverty, but rather a distribution issue, not an output issue. The total output of such an economy could be extremely large, which in itself is a scenario of abundance. The real disaster, if it occurs, depends entirely on who owns the capital stock that produces this abundance.

The most pessimistic scenario secretly assumes two conditions: humanity no longer retains any market-priced advantages and does not own any productive assets. But neither of these points is a natural law.

The services provided by humans may gain a premium from care, status, authenticity, and the demand for "hope that humans will serve." As overall abundance increases, this space may expand rather than shrink.

Wherever intelligent agents must interact with the physical world, bear legal responsibilities, or enter regulated fields to perform tasks, cost baselines still exist, and human labor may concentrate in these areas.

Meanwhile, if displaced workers possess capital through pensions, widespread stock ownership, or distributed tokens, then the losses from the decline in labor share can be offset by their participation in capital gains.

The core conclusion here must be stated directly: the labor issue and the ownership issue are essentially the same issue.

A decline in labor share will only evolve into a social disaster under conditions of highly concentrated ownership. If ownership is sufficiently broad, the same automation will become a widely shared abundance. The fear of large-scale technological unemployment, upon further analysis, is actually a fear of the distribution of capital.

Therefore, the ultimate answer does not lie in preserving all old jobs, but in reconstructing ownership. This also reconciles the tension presented at the beginning: individuals will indeed be significantly amplified by intelligent agent tools, while the share of labor in total output may continue to decline. Both can coexist. The same person may become far more powerful than in the past, yet only receive a smaller proportion of total output.

This contradiction can only be truly resolved at the level of ownership. An enhanced individual who owns a share of machine assets can share in the abundance created by machines; an enhanced individual who owns nothing remains just a more efficient worker, whose wages are still suppressed by the declining costs of surrounding intelligent agents.

This makes the concentration of power a decisive issue. "Concentration is the default" is not an absolute physical law. In open protocols, forkable systems, and the history of commodification, there are also many cases of power being decentralized rather than concentrated. The real question to judge is: at which layers will concentration prevail?

The answer lies at the layer where two conditions overlap: strongly increasing returns and non-forkable bottlenecks.

Network effects and data flywheels will make each new user increase the value of the product for the next user, while non-forkable bottlenecks will prevent competitors from replicating core advantages.

Open-source code can be forked, but dominant currencies, regulatory licenses, deep liquidity pools, and a key intervention that can reverse system outcomes cannot be easily forked. Conversely, at layers where cutting-edge capabilities are rapidly commodified, switching costs are low, and forking is credible, openness can truly decentralize power.

Thus, the correct question is not whether the economy of intelligent agents will concentrate, but which layers sit on non-forkable bottlenecks.

The foundational model layer, often seen as the greatest barrier, is actually the most likely to gradually commodify. Cutting-edge capabilities are often caught up by open-source weights within a year or two, and reasoning prices have already dropped by several orders of magnitude.

Persistent rents are moving away from bare models and migrating to complementary products of models: proprietary data, distribution channels, and a truly "user-understanding" intelligent agent forming a lock-in effect. The bridge layer is more about security than rent issues, as it has long been the most vulnerable surface in the entire field. Oracles may concentrate on a few data sources that everyone can use, but still have some degree of substitutability.

The truly lasting barriers are likely to be the identity layer and the intervention key. The identity layer has a winner-takes-all dynamic, as a verified identity can be reused across multiple scenarios; the intervention key has ultimate control, as those who can reverse deterministic core outcomes control the entire company in extreme cases.

Meanwhile, token governance, packaged as a democratizing tool, will slide towards oligarchy unless the system is deliberately designed in reverse, especially if it still adopts a one-token-one-vote structure. The concentration issue of stablecoin issuance layers also needs to be directly addressed, as this structure is directly related to the industry in which Circle operates.

A dominant stablecoin issuer can gain the returns generated by the reserve assets behind the currency it intermediates. The stablecoin issuer and its industry are the parties that may benefit from this concentration. The issuer's reserve returns are largely shaped by policy.

It exists in its current form because laws often prohibit issuers from directly paying reserve returns as interest to holders, but increasingly allow value to flow back to the ecosystem through usage-based incentives or through distribution partners.

The structure created by statutory law can also be redistributed by statutory law. Returns that cannot be paid as interest can be redirected through participation mechanisms: through competition, regulation, or directly sharing with those who hold and use this currency.

Widely distributing reserve income to the ecosystem has become a practical way of operation in the industry, and the proportion is still increasing. This portion of rent is competitive and also relies on policy. Reasonable policies are likely to require more value to ultimately flow to holders and users.

However, the sharpest concentration risks may not necessarily exist at the stablecoin issuance layer, but may exist at the smart infrastructure layer, especially for entities that control both economic infrastructure and smart infrastructure.

This has gradually entered policy discussions: should society as a whole hold part of the capital equity of the largest AI companies? The concentration issue is also a geopolitical issue, as the layers that can gather rents are also the most likely to become weapons.

Economic operating systems should remain technologically neutral, governed by multiple stakeholders rather than controlled by a single nation. At the protocol and technical level, this neutrality can be achieved and is worth defending, as these layers can be forked and do not have a single owner.

But as long as a certain country issues a dominant currency and can freeze, confiscate, or exclude specific entities through the track, the currency layer cannot achieve complete neutrality solely through governance design. A dominant digital dollar may be the most realistic path for the entire system to achieve global scale, but it is also the greatest single control concentration risk. Claiming that the entire stack can remain neutral is an overly naive narrative.

A more credible statement is to clearly distinguish which layers can remain neutral and which layers cannot truly be neutral. Entities that control settlement tracks, reserve currencies, or identity layers will simultaneously gain a panoramic monitor and a control switch. This is what can already be seen in the "weaponized interdependence" of traditional financial sanctions and clearing exclusion systems.

Deep economic interdependence does indeed marginally increase the cost of conflict and may reduce the probability of conflict between some countries, especially among democratic nations. The economy of intelligent agents may become the most favorable economic order for peace in history, or it may become the easiest economic order to weaponize. Which one it ultimately becomes depends on whether key barriers can be neutralized or will be seized by a few powers.

So, what does the promise of abundance mean?

The economy of intelligent agents may indeed become the largest poverty reduction opportunity in history, and it may significantly popularize health, education, and information services. But abundance and inequality can very well rise simultaneously.

The strongest optimistic argument comes from consumer surplus. If intelligent agents drive the marginal costs of medical diagnosis, personalized education, legal consulting, and expert knowledge towards zero, a person with very little monetary income may still see a significant improvement in their actual living standards, even if the income and wealth gaps on paper expand simultaneously.

These two sets of accounts may develop in different directions. A society can be more unequal in terms of income and ownership while being more equal in what people can actually consume and accomplish. Ultimately, which side dominates does not depend on emotions or prophecies, but on the breadth of ownership and the design of fiscal policy.

Broad ownership combined with redistribution mechanisms can turn the same automation into a story of shared abundance; concentrated ownership combined with a passive, hands-off state will make it a social upheaval built on extreme wealth concentration.

The decline in labor share, the concentration of power at non-forkable bottlenecks, and the geopolitical risks brought by weaponizable infrastructure are not three independent issues, but the same issue: who will own and govern the layers that are concentrating value?

At these layers, concentration is the default result and is reinforced by the deep laws of platform economics. Shared abundance can still be realized, but it must be actively constructed against the gravitational pull of capital. The key tools are ownership distribution and governance of key barriers.

Technology has for the first time made better outcomes shift from "imaginable" to "buildable," but it will not make this outcome happen automatically. How ownership is distributed broadly enough and how key barriers are governed rather than seized are the last parts that need to be answered.

IX. Civic Vision

The economy of intelligent agents may sever the originally stable link between labor and output share.

The real answer is not to desperately defend all jobs, but to broaden the ownership of the capital that is accumulating value, including intelligent agents, models, infrastructure, and companies. This is the answer that the entire argument ultimately points to, and it is not unattainable.

The same set of on-chain architectures, if left to default trends, will concentrate ownership at a few key barriers; if designed differently, it can distribute ownership, returns, and governance more broadly than any economic order in history.

First, we need to return to the historical lineage of joint-stock companies, as it marks the scale of possibilities.

Joint-stock companies are a profound social technology. They allow strangers to pool capital, share the long-term results of a business, and thus give rise to a complete set of systems such as capital markets, corporate entities, modern banks, and exchanges. They also expanded the scope of participation in business development from a few merchants and monarchs to a broader population.

The on-chain economy continues this lineage and can, in principle, further transcend it. For the first time in history, the relevant technological conditions are in place: not only capital itself, but also corporate governance and upward returns can be distributed to an extremely large number of people at nearly zero management cost.

Digital tokens are an important component of this. This is also the goal that the Web3 movement attempts to express through "read, write, own": platform users not only provide the attention and data that the platform monetizes but can also share in the governance rights and economic benefits of the platform.

This ambition is not new; the real change lies in the collapse of execution costs. There is an appealing argument that early broad ownership movements, such as guild socialism, cooperative movements, and Kilt socialism, failed simply due to the lack of coordination tools provided by today's on-chain tracks; their ideals were correct, just waiting for the infrastructure to mature.

However, this argument overstates the role of tools and does not align with history. These movements did not merely fail due to logistics and coordination costs but were defeated by power: capital, employers, and the state retaliated through repression, hostile legislation, and even violence.

Successful cases actually prove the same point. The Mondragon Cooperative has tens of thousands of worker-owners, and credit unions, mutual insurance organizations, and large consumer cooperatives have achieved distributed ownership and governance at a real scale for decades. They relied on sound legal forms and enduring institutions, not blockchain.

Meanwhile, in the evolving state-led market capitalism system, the trend may not necessarily lead to a deeper democratization of industries; it could evolve into a new form of stakeholder participation: allowing society at large to hold those technological platforms that are amassing the most capital and are the most disruptive.

Thus, the claims of the on-chain economy need to be appropriately narrowed, which ironically makes them more robust. On-chain tracks can reduce the coordination and transaction costs of broad ownership and dismantle the gatekeepers that have historically stood between ordinary people and a share of capital. This is a real and significant contribution.

However, on-chain technology itself cannot eliminate the power asymmetries that truly defeated historical movements. Technology can make distribution cheaper and more feasible, but it will not automatically lead to distribution.

What truly makes it happen is still politics.

Moreover, in the absence of special designs, the default outcome may still be re-concentration.

The actual records of token ownership have provided clear warnings: pre-sales and insider allocations, team unlocks, especially freely tradable tokens that flow back to the largest holders through secondary markets after gaining value, have made on-chain ownership at least as concentrated as traditional equity, and sometimes even more so.

One token, one vote governance, by design, is oligarchic.

Liquidity itself may become the enemy of broad ownership.

Therefore, while "this architecture can be distributed" is not wrong, it is almost a hollow statement. The real question is whether the distribution mechanism is written into the system against the gravitational pull of concentration. This design is achievable, but it must specify concrete mechanisms and accept corresponding costs.

Ownership can be distributed based on participation and contribution rather than merely purchasing power, allowing those who actually use and build the system to gradually own it. It can also weaken the re-concentration caused by the secondary market through vesting periods and transfer restrictions, setting progressive caps on individual allocations.

These measures all combat the natural gravity of capital and will incur real costs. There is also a deeper trap often overlooked by optimistic narratives: distributing ownership does not equal distributing power.

Broad and fragmented shareholding can exist alongside ordinary holders having almost no governance influence. Retail shareholders in traditional public companies have already proven this.

In on-chain systems, this gap may even be larger, for instance, through founder super voting rights, foundation control, low-participation delegated governance, and whale-dominated voting. More critically, enduring power often resides in controlling checkpoints, not just in cash flow rights.

The intervention keys that can reverse the system and the identity layers that are repeatedly used across multiple scenarios may be the true centers of power. Even if economic benefits are distributed to a billion individuals, those who hold the intervention keys may still control the entire company.

Therefore, distributing governance rights is a task that is independent of distributing economic benefits and is even more challenging, and it must directly target these control points.

The debate over whether state power should intervene in the deployment of foundational models and security controls is a complete reflection of this issue. This means that governance designs need to be established that can sever the direct link between wealth and votes.

For example, one-person-one-vote guaranteed by personality proof, quadratic voting, belief voting, and rights weighted according to real contributions. At the same time, public interest governance must be implemented for critical system infrastructure. Independent directors or public interest director seats, golden shares, and custodial, multi-party authorization, and complete traceability controls for intervention keys and identity layers can be set up.

This scope should also include the most powerful, capital-intensive AI foundational models, ensuring that these infrastructures are no longer merely switches controlled by private entities. If only upward benefits are distributed while control checkpoints remain concentrated, it is merely an incomplete solution. It is essential to distinguish between the relatively easy half of the citizen vision and the truly difficult half.

The market's inherent distribution mechanisms: tokens, universal access, and long-tail participation will expand the overall market for stablecoins and on-chain finance, but relying solely on these mechanisms is far from sufficient.

The ultimate outcome depends on the breadth of ownership and fiscal policy design. Fiscal policy cannot be quietly omitted, as this part requires the state to directly confront interest groups benefiting from concentrated structures.

The truly viable position should be "and," not "replace public policy with market mechanisms." On one hand, broaden ownership through institutional design; on the other hand, implement progressive taxation on capital and automation benefits, incorporating public goods that a prosperous society should disseminate into public supply.

In a stronger form, a public capital equity can be established: setting citizen dividends or sovereign rights for critical infrastructure, allowing the public to directly share the capital gains that are replacing labor income. Market mechanisms are responsible for distributing what the market is willing to allocate, while fiscal systems are responsible for distributing what the market refuses to allocate actively.

In the layers of identity, settlement, and model access, open, forkable, and mandatory interoperable standards should be adhered to, ensuring that critical checkpoints are always competitive in a technical sense, and allowing captured participants to form constraints through "exit."

For critical system infrastructure, public interest governance and custodial control should be implemented. Ownership can be distributed based on contribution and usage, and transfer restrictions can be used to reduce re-concentration.

On this basis, a fiscal tool can also be added: taxing automation or capital gains to establish a broad on-chain ownership dividend. The objects of distribution should not only be transfer payments but can also include shares of productive capital itself. These are just starting points, not a completed program.

However, the so-called "executable building blocks" must at least genuinely include these mechanisms, and many of these measures will require existing stakeholders to pay real costs. None of this will automatically win just because it is logically correct; the underlying political economy must be confronted. Existing participants benefiting from concentrated structures are often precisely those most capable of influencing rule-making.

A path inclined towards broad distribution will not naturally prevail because it is more just; it can only win through real checks and balances. These checks and balances include: defeating technological capture through forkability and open standards, directly constraining control layers through antitrust and public obligations, implementing public ownership or governance for the most critical tracks, and establishing a broad owner class with real economic interests willing to defend the relevant systems.

This owner class itself is the voter base that sustains distributional politics. "This cannot be solved solely by one country" is correct, but merely calling for civil society, businesses, and political leaders to form a common vision does not automatically complete the transition.

This vision must be fought for in real politics, taken from the hands of vested interests.

Beneath all economic issues lies a more fundamental question. If labor is no longer the primary organizing principle by which people gain social position and voice, then ownership may have to take its place. A share of equity in the capital stock may become the new economic foundation of citizenship, providing the footing, security, and social participation qualifications that past wages provided.

As machines absorb more and more work, those areas that always belong to humans—care, creation, craftsmanship, judgment, value produced by human-made or service-based efforts, community, and meaning—should not be viewed as economic surplus.

They may very well be the things that a prosperous society finally has the capacity to truly cherish. The premise is that society is willing to price these activities and grant them dignity, rather than viewing them as the remnants left after machines complete the rest of the work.

In such a world, how people should live, how they should spend their time, and how they should be recognized is the most profound open question posed by this economic model. This architecture does not provide direct answers but creates possibilities for better answers. Which answer is chosen depends on us.

This is the entire argument, which ultimately returns to the starting point. The agency economy is the on-chain economy: artificial intelligence provides labor, while the on-chain foundation provides the forms through which currency, decision-making, coordination, and ownership can be expressed.

This architecture has its pros and cons, with its default mode being centralization, concentrating income, power, and all the infrastructures that can be weaponized. But we also have, for the first time, tools that can widely distribute ownership, outcomes, and governance rights, allowing the same mechanisms to shift towards shared prosperity.

Disclaimer: This content is provided for general branding and informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online events, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets or to use any services. Crypto assets are highly volatile and may result in loss. WEEX services and online events may not be available in all regions and are subject to applicable laws, regulations, and eligibility requirements. You are responsible for ensuring that your use of WEEX services complies with local laws and for carefully assessing the risks before participating in any crypto-related activities.