Understanding Circle Founder Jeremy Allaire's Paper on the 'Agent Economy': Insights into How Economic Structures Will Transform in the Next Decade

Original text by Jeremy Allaire, founder of Circle

Compiled by: Odaily Planet Daily, Qin Xiaofeng ( @QinXiaofeng 888 )

Editor’s Note: On July 13, Circle founder Jeremy Allaire published a research paper titled "The Agent Economy," exploring the integration trends of AI Agents and future economic systems. Allaire stated that as AI Agents begin to take on corporate tasks and value circulates natively through open, programmable networks, the Agentic Economy and Onchain Economy will ultimately become two sides of the same economic system.

"This paper is the culmination of decades of building internet infrastructure and crystallizes a question I have focused on from the beginning: that open software and open networks can not only change how we share information but also reshape our social, political, and economic landscapes. Many of the ideas in this paper stem from two core beliefs I had when founding Circle. First, that money can flow through open protocols just as information does on the open internet. Second, that blockchain is a network computer: it is a foundational platform where autonomous software and machines can store value, exchange value, and coordinate economic activities directly without human intervention," Allaire introduced his research intentions.

He added that these initial concepts have evolved over time, leading to a deeper understanding of how financial and economic systems can integrate with software and the internet. With the emergence of this integration alongside truly powerful artificial intelligence and agent systems, this theory has further expanded: it describes not only a new type of currency or network but also a completely new economic operating model and its implications for humanity, labor, capital, ownership, and new social contracts. This is precisely what the book aims to explore.

---

01 The Convergence and Deconstruction of Enterprises

Every major transformation in the internet era follows the same path: it does not stem from a single invention but rather from multiple technologies maturing and suddenly converging. The internet, mobile, cloud, and social media are all examples of such convergence, repeating the same underlying patterns.

The Law of Convergence

When various capabilities converge, the costs of once-expensive activities approach zero, and once costs reach zero, the scale of that activity can explode. This is true for information on the internet, communication through mobile and social media, and software in the cloud.

Today, two new systems are converging, directing the same forces toward two areas that the internet has never fully digitized: intelligence itself and the economy itself. The first is intelligent systems, composed of AI models and the agents built upon them, which drive the costs of thinking and working toward zero. The second is the economic system, constituted by blockchain, where money, contracts, and coordination operate in software form, driving transaction costs toward zero. The two empower each other, and the core assertion of the entire treatise is: these are not two parallel trends but two sides of the same economic entity.

Two Operating Systems

The intelligent system is the most critical because it changes the nature of software.

You no longer program; instead, you issue commands in natural language, and it infers the answers rather than following fixed steps. Its basic unit is the Agent: a reasoning process to which you delegate tasks. This transforms software from a program executed word-for-word by machines into work that you can delegate to thinking machines, allowing the core tasks of enterprises to be broken down and reconstructed into skills that agents can perform.

Beneath brands and buildings, companies are essentially organized thinking: products, marketing, sales, finance, legal, plus the external companies they hire. These are almost all human labor, and human labor is the largest cost in the economy, which is precisely the target of cheap and powerful intelligence.

Decomposing Enterprises

It also disrupts the traditional explanations of why enterprises exist. Enterprises grow because the costs of coordinating external work are high, thus internalizing it; however, when any non-physical work can be completed by agents that you can instantly find, hire, and pay, this logic weakens, allowing one person to accomplish work that previously required an entire department.

It first descends upon software and other information-intensive work, while in the physical realm, it is slower and still awaits breakthroughs in robotics. This is not merely about personnel reduction: a person working alongside powerful agents will become extremely efficient, while judgment, interpersonal relationships, and ultimate responsibility remain with humans. This leaves a tension that needs further exploration, which will be addressed later through ownership: even as the proportion of the economy paying for human labor decreases, individual capabilities can be amplified.

02 Assembling, Coordinating, and Why Enterprises Should Go Onchain

Once enterprises are decomposed into various skills, the real question is no longer which can be automated, but how these fragments can be reassembled.

The answer is the orchestration layer: a general manager agent receives objectives, breaks them down into tasks, assigns them to specialized agents, and then stitches the results together, with supporting software passing context and memory between each step. The same mechanism applies to any function, so marketing, finance, sales, and product are essentially the same set of machines applied to different tasks.

Humans will not disappear. Some will remain within the closed loop, executing or verifying tasks that require human judgment. Others will rise above the closed loop, setting goals, defining standards, monitoring quality, and deciding when machines should stop and consult. The shift from executing work to supervising work is the true form of human oversight, and the corresponding tools are on the way.

Orchestration Layer

When a company clarifies a task sufficiently for internal operation, it is also clear enough for external hiring, thus an open agent market almost forms as a byproduct.

This market may take two paths. It may evolve into a few large platforms selling agents like utilities, or more interestingly, it may form a true labor market composed of specialized agents, as deep expertise remains valuable, and enduring enterprises will be those whose agents delve deeply into a particular field.

However, hiring software that can be assembled anywhere in the world requires trust, which is precisely the issue that pushes all links onto the chain.

The solution is identity layering. At the base is a public blockchain that anyone can verify. On top of that is real-world identity verification, similar to the verifications that banks have been running on a large scale, agent wallets and credentials, and reputations that accumulate over time but are tied to verified real creators. Together, these form a chain of accountability: every action of an agent can be traced back to the real individual or company responsible for it.

Integrity First, Accountability Throughout

A single company’s private database cannot achieve this because trust locked within a single operator cannot be transmitted, while identities rooted in public chains and real-world verification can. Therefore, autonomy here does not mean anonymity. Behind an autonomously acting agent, there is always a person responsible for it.

Chain of Accountability

03 Monetary Foundations: Speed, Security, and Finality

Agents need currency that can be held and transferred to operate at machine speed, whether in large or small amounts, without needing to stop and verify the reliability of the currency itself with each payment. The last point is crucial; it points to a traditional answer: fully backed, final settlement currency operating on an open network.

Speed Replaces Leverage

Let’s start with speed because it will reorganize everything else.

When the cost of transferring currency approaches zero, settlements are completed in an instant, and currency can be controlled by software, the same dollar can be reused multiple times in a short period, and any amount can be utilized the moment it arrives, making small payments between agents finally feasible. This is precisely the pattern that information and software have already followed on the internet, now extending to currency.

Each part of the answer has its reason for existence.

A natural rebuttal is that banks create speed by repeatedly lending the same deposit, so will full backing stifle credit? It won’t: when the turnover speed of currency is fast enough, a dollar can be locked for a few seconds and then lent out, thus speed serves the role that leverage once did, and credit is rebuilt on the foundation rather than eliminated.

Why Base Currency Should Not Bear Risk

Why insist that base currency cannot bear any risk? Because the danger of risky currency is proportional to its turnover speed. What used to take weeks to happen in a bank run can now occur in minutes, and agents with instant settlement cannot stop to judge whether each dollar is reliable.

Fully backed currency is the only currency that is worth exactly one dollar to everyone, everywhere, without relying on a national safety net that cannot cover a global system. Settlements must also be equally certain: not a potential end after a period of time, but an end within a second, settlement is settlement.

Institutional Framework

Refunds and fraud protection still exist, but are built as optional layers on top, such as escrow, refund pools, and insurance, rather than being embedded in the currency itself. These safeguards do not activate automatically; they rely on real institutions being built, large issuers being regulated, isolated from bankruptcy, and backed by increasingly secure reserves.

There is a boundary that must be clear: Holding currency does not generate any yield. Reserve yields belong to the issuer and flow into the ecosystem, but when you pursue yield, you no longer hold currency; you are lending currency and taking on risk. Confusing the two will undermine the entire security argument.

04 Credit Markets: Machine Underwriting, Agent Working Capital, and Prudent Regulatory Layers

When base currency is fully backed, credit does not disappear; it shifts to the other side of that line and returns with a stronger posture, covering a broader population, pricing more accurately, and failing more visibly than the system it replaces.

The long-tail effect under underwriting constraints

The key is to redefine the problem. A large number of borrowers, including small merchants, gig workers, families, and now agents, are underserved not because they are high-risk, but because the cost of reviewing each small loan exceeds the value of the loan itself. Credit allocation depends on underwriting costs, not the quality of the borrower. Reducing underwriting costs can enable a large number of previously creditworthy but overlooked borrowers to receive service.

Data flywheel

What drives cost reduction is a data flywheel: On-chain activities are structured, verifiable, and real-time, which makes risk models far superior to the previously scattered records; better data leads to better loans, which in turn attracts more activity and more data.

People naturally worry that this will record everyone's financial situation on a public ledger, to which the answer is simple: Being on-chain does not mean being public. New privacy technologies allow individuals to prove the information that lenders need, such as their credit status or loan balance, without disclosing specific details.

On-chain does not mean public

The core is a truly new type of loan: Working capital provided for agents. It has an unusual predictability because it removes the biggest variable in human lending—whether the borrower is willing to repay—simplifying risk to a short-cycle, limited-scope question about specific work content.

Agent working capital

Imagine an agent borrowing four dollars of computational resources to complete a ten-dollar job they have been hired for. The lender is not guessing character; it is simply pricing the probability of the work being accepted. Collateral disrupts the conventional model: instead of slowly seizing unrelated assets through the courts, the loan is first secured by the payment for the work itself, automatically claiming it, and backed by the deposit made by the agent, their reputation, and ultimately the real individuals behind them.

The result is credit that is cheaper, more widespread, and simultaneously safer, which seems impossible until you understand that the yield comes from better information, not more lending.

The honesty required for this claim is that this predictability will diminish over time: tasks completed in seconds are nearly mechanical, while financing over months will revert to ordinary risk levels.

Thus, machine credit will not replace human credit; it becomes a new low-risk benchmark, with human loans priced against it.

Moreover, all of this is under surveillance: risks will manifest as they accumulate, and automatic brake mechanisms will steadily increase the costs of flooding into the same model or provider, in addition, insurance costs are also priced based on actual conditions rather than outdated averages.

05 Natural Globalization

The architecture has exactly three layers.

The bottom layer is currency: stablecoins as units of account and final settlement means. The middle layer is the economic operating system: coordination, contracts, and value exchange operate in the form of programmable smart contracts and have final settlement capability. The top layer is the agent execution layer: actual work is completed here, driven by AI and the cloud.

The key aspect of these three is their position of existence. Each is software, each operates on the internet. Each also replaces things that were once tied to the nation: software currency replaces the national banking system stitched together by slow agents; the middle layer moves contract execution from national courts to code that operates the same anywhere; agent execution replaces local labor with work that has no homeland.

Therefore, the economy built on these layers is inherently borderless. This is the meaning of "natural globalization": not an added feature, but an inherent property of its constituent materials. Throughout history, economic activity was primarily national, and cross-border required extra effort; now, economic activity is primarily global, and the framework of the nation needs to be added afterward.

No single indigenous jurisdiction

An economy without a homeland does not escape the law; it will be subject to too many laws at once, with conflicting rules from many jurisdictions, yet no single location to determine which set applies. The solution is to shift the question from "where something happens" to "who is behind it," regulating each agent traced back to an accountable entity, while the country where the user actually resides sets the conditions for market access.

Execution moves to the edge, where currency and identity cross between the open world, the regulated world, and the private world, checking before payment settlement rather than reporting after clearing. This does not require everyone's financial public ledger: by default, disclosures remain private and are only shared with permission.

A healthy system also retains a truly private space, namely a digital version of cash, so control lies at the regulated edge, not the core. The most powerful tool—the ability to freeze or withdraw funds—is only legal under true due process: documented, time-limited, requiring multiple parties' involvement, and allowing for appeals.

Multi-currency and intangible foreign exchange

Currency exchange also becomes intangible, as with each major currency on-chain, you hold your local currency, the other party receives their local currency, and the exchange is completed at the underlying optimal rate. Sovereignty is reshaped, not lost: a neutral network allows a country to issue its currency on the same track rather than relying on others' currencies.

The real danger lies in the transition period, not the endpoint, as people can escape weak currencies faster than ever, thus it must be managed.

This economy has both equalizing and centralizing tendencies, with centralization being the default state, and broad sharing being a more difficult, buildable alternative. The same machine can both enforce accountability and implement censorship, with the choice in our hands.

06 Supply Side: From Subscription to Consumption

The agent economy requires a supply side, which consists of services that agents can call upon, hire, and pay, forming in two waves.

First, existing software and data encapsulate themselves so that machines can use them, with pricing aimed at agents rather than individuals. Second, new specialized agents are constructed, delving into specific fields and selling their work. The deep transformation lies in the pricing method: Value shifts from access to work outcomes, resetting the software business.

For thirty years, software has been sold by seat, charging periodic fees to logged-in individuals. But now the customer is the agent executing tasks, thus purchasing the work itself rather than login access. Seats as a unit of charge are disappearing, although subscription models will not vanish; pricing is reshaped around new units of work in various forms, from pay-as-you-go to committed budgets, to pricing based on deliverables.

The same logic extends down another layer, where the flow of funds resides.

With the surge of specialized agents, buyers purchase outcomes from agents rather than raw outputs from models, and agents will select from competitive models to complete work at the lowest cost within quality allowances.

Models are costs, agents are business

This is already happening: tools that route each request to the best model have become essential within a year, as the price gap between models is so vast that using expensive models for simple tasks is sheer waste. Thus, models become cost items, and agents become the business itself, with value flowing to the party that has the customers, context, and outcome responsibility.

This is a trend, not a law, as the best model creators maintain true pricing power on the most difficult tasks and can move up into the agent layer themselves; the possible outcome is a dumbbell structure, with a large middle section commoditized while frontier areas retain value.

The era of micro-payments for labor has arrived

Beneath this, an old dream finally comes true: Micro-payments. They have never succeeded on the consumer internet, partly because settlement is expensive, but mainly because people hate deciding whether each little thing is worth a penny.

Machines do not have this hesitation, and settlement is now almost free, so micro-payments finally arrive, not for content, but for small unit work between agents.

The optimistic narrative overlooks one issue: if agents can hire other agents and tools themselves, expenditures can spiral out of control quickly, thus the economy needs a spending control layer, including caps, budgets, and approvals, which itself becomes a product category, refining the overall vision rather than weakening it.

07 On-Chain Companies

As agents take on more work for businesses, the businesses themselves also need a new habitat.

A company where work is completed by agents holding currency, signing contracts, and operating around the clock needs a place where all this can truly happen: currency flows programmatically, rules operate in software, and external transactions settle at machine speed. That place is the on-chain economy.

Two parallel paths

Thus, agency companies and on-chain companies are actually two sides of the same coin: one side describes who is doing the work, while the other describes the form the work takes. This is the core of the entire thesis: an economy run by software agents must operate on software money, software contracts, and software governance; otherwise, it cannot function at all.

This does not mean— and this distinction is more important than any other— that every company dissolves into a collective run by tokens.

The future is a hybrid, moving along two tracks.

On one hand, existing companies gradually put their shares and governance on-chain while retaining their familiar legal forms, a slow change driven by the most cautious institutions in the financial sector.

On the other hand, new, highly agent-based companies are building on-chain from day one, pulling everyone else forward. Even these new companies will not escape the law simply because they are born from software: the existence of law and limited liability comes from the government, not from lines of code, so they still need to wrap themselves in a thin legal shell. What flips is the ratio: the legal shell becomes thinner, while the on-chain working entities become thicker.

Even De Novo needs a shell

Two admonitions keep this honest. First, a shared ledger can prove what happened, in what order, and by whom, which is a real advance, but it cannot prove that an action was authorized, wise, or loyal; a perfect self-serving transaction record is still a self-serving transaction. The ledger is a better witness, not a better conscience, so responsibility still falls on the humans who design the agents and are supposed to supervise them.

Second, contracts become programs in how they are executed, running automatically in common, clear cases, but remain legal documents in how they are adjudicated, as code runs verbatim while law leaves room for intent, error, and fraud.

The best way to envision this is with a core that is reliable, edges judged by humans, and a few contentious cases handled by external data sources, arbitration, and shared, time-limited, recorded overwrite mechanisms, because ultimately, whoever holds the overwrite power holds the company.

08 Impact and Concentration of Power

The agency economy holds both the greatest opportunities and the most severe risks of this era in the same hand; they are not two futures to choose from but rather the joint result of the same machine, whose balance is still undetermined.

Starting with labor, it must be stated with sufficient caution to withstand the oldest disputes in economics. The claim is not that automation overall destroys jobs— this assumption has been proven wrong for two centuries. The real issue is the proportion of national income that human labor occupies and the wage levels that human work can command. People may still find employment in the weakest tasks of machines, but the pay for those jobs may drop to levels insufficient to support a family, which on paper is full employment but in practice is a crisis.

Labor share, not employment

The conditions for this situation to hold are: software takes on new tasks faster than people can retrain; the price of agent labor continues to decrease with computing costs, pulling down wages; and— a breakthrough that is truly different from all past waves— capital can self-finance growth, with agents earning money to build more agents. A loom has never earned enough to buy the next loom; agents can.

Capital → Software → Capital Cycle

Two candid admonitions prevent this from descending into fatalism. Even if all the above conditions hold, the result is still a distribution problem, not a scarcity problem, because output can be enormous— this is the theory of abundance. Moreover, the pessimistic view secretly assumes that humans no longer have an advantage and own nothing,

Neither of these points is predetermined: human work may enjoy a premium in care, status, and authenticity, and if displaced workers own capital, the declining labor share can be offset by the capital share they participate in.

This is the crux: it should be stated clearly: the labor issue and the ownership issue are the same issue. A declining labor share is only a disaster when ownership is concentrated; if ownership is widely distributed, then the same automation is merely shared abundance. This makes concentration a decisive issue, worthy of analysis rather than assertion.

Labor and ownership are the same issue

Concentration is not a natural law; open standards and forks have a long history of decentralizing power. It only prevails when strong network effects encounter non-forkable bottlenecks: you can copy open-source code, but you cannot fork the dominant currency, licenses, deep liquidity pools, or overwrite keys.

The places where power is most likely to concentrate are not AI models— they tend to commoditize— but rather identity layers, overwrite rights, and dominant currency issuers, who earn the returns on the currencies they handle. The author is in the last of these fields and admits this, and he presents a view that goes against his own interests: that return is a policy choice, and what policy creates, policy can also redistribute.

The same control points both aggregate profits and can become weapons; history is a warning, so those dense connections that raise the cost of conflict can also become tools of conflict. How it goes depends on whether these control points remain open or are captured.

09 Citizen Vision

If the agency economy breaks the link between labor and output shares, the answer is not to defend old jobs but to expand ownership of the capital that is capturing value— agents, models, infrastructure, and companies. The same architecture, if left unchecked, will concentrate at a few control points, but it can also disperse ownership, returns, and governance more widely than any previous system.

Expand ownership, not defend jobs

Inheritance determines scale: joint-stock companies once allowed strangers to pool funds and share in the success of enterprises, extending participation beyond the wealthy and royalty. The on-chain economy can extend this further, as it now has the tools to grant not only ownership but also governance rights and upward mobility to a large number of users at almost zero administrative cost.

This idea is not new; what is fresh is that the cost of putting it into action has become low. But capability does not equal outcome, and this section holds itself to strict standards: listing the mechanisms that truly work, including those that cost the author himself.

Real history refuses to gloss over. Early movements for widespread ownership did not fail due to the paperwork issues that blockchain now solves, but rather fell to power.

On-chain mechanisms lower the cost of shared ownership and remove some gatekeepers, which is true, but they do nothing to address the power imbalances that truly stifled these movements.

Worse still, the default setting is a re-concentration of power: internal distributions, especially in open secondary markets, will pull tokens back to the largest holders once they have value; and the "one token, one vote" system is designed from the outset to ensure the rich dominate. Liquidity ultimately becomes the enemy of widespread ownership.

Thus, ownership must be obtained through participation, limiting transfers, and setting caps, while designing shared mechanisms that consider this tug-of-war, and also accepting the reality that liquidity and breadth cannot be maximized simultaneously.

Moreover, there is a deeper trap: shared ownership does not equal shared power. You can allow a billion people to participate in economic activity, but those who hold the final decision-making power still control the company. Therefore, decentralizing governance is an independent and arduous task, aimed directly at these control points.

Its stance is: design to expand ownership and combine it with fair capital and automation taxes, public supplies that should disseminate abundant resources, and sharing public interests, while ensuring that the public can share in the value created by these infrastructures. The most explicit standard the author uses to measure his own interests is the returns on stablecoin reserves: this is a policy product that should lower returns through competition and ultimately return to those who hold these funds, including the issuers associated with him.

None of this can succeed solely on its own merits, as the beneficiaries are the rule-makers, so counterbalancing forces are needed: open standards make rent-seeking no longer an obstacle, impose public directives on the control layer, and a broad base of true stakeholders to defend their rights.

All of this ultimately leads to a core question: if labor is no longer the path for people to gain status and voice, then ownership may have to take its place. Infrastructure is not a fate. Whether this will become the most balanced economy in history or the most concentrated economy is not a prophecy that can wait, but a design problem to be solved and a political struggle to be won. The test of whether we are sincere is whether we will first constrain ourselves.