DWF In-Depth Report: AI Outperforms Humans in Yield Farming Optimization in DeFi, But Complex Transactions Still Lag Behind 5x

Original Article Title: Will Agents take over DeFi?
Original Source: DWF Ventures

Translation by: DeepFlow Tech

Key Points

Currently, automation and agent activity account for about 19% of all on-chain activity, but true end-to-end autonomy has not yet been achieved.

In narrow, well-defined use cases such as yield optimization, agents have shown better performance than humans and bots. However, for actions involving multiple aspects like trading, humans outperform agents.

Among agents, model selection and risk management have the greatest impact on trading performance.

As agents are adopted on a large scale, there are several risks related to trust and execution, including sandwich attacks, strategy congestion, and privacy trade-offs.

Continued Growth of Agent Activity

Over the past year, agent activity has steadily increased, with both transaction volume and count on the rise. We have seen significant developments led by Coinbase's x402 protocol, with players like Visa, Stripe, and Google joining in to introduce their own standards. Most of the infrastructure currently being built is aimed at serving two types of scenarios: channels between agents or agent calls triggered by humans.

While stablecoin transactions have gained widespread support, the current infrastructure still relies on traditional payment gateways as the underlying layer, meaning it still depends on centralized counterparties. Therefore, the "fully autonomous" end state where agents can self-fund, self-execute, and continuously optimize based on changing conditions has not yet been achieved.

Agents are not entirely unfamiliar to DeFi. For years, there has been automation through bots in on-chain protocols, capturing MEV or achieving outsized returns that would not be possible without code. These systems have operated very well under clearly defined parameters that do not change frequently or require additional supervision.

However, the market has become increasingly complex over time. This is where we see the entry of the new generation of agents, with the past few months on-chain serving as an experimental ground for such activities.

Agent Performance in Action

According to reports, agent activity has seen exponential growth, with over 17,000 agents launched since 2025. The total volume of automation/agent activity is estimated to cover over 19% of all on-chain activities. This is not surprising as over 76% of stablecoin transfer volume is estimated to be generated by bots. This indicates significant growth potential for agent activity in DeFi.

Agents exhibit a wide range of autonomy, from chatbot-like experiences requiring high human supervision to agents that can formulate market-adaptive strategies based on target inputs. Compared to bots, agents have several key advantages, including the ability to respond and execute new information within milliseconds and expand coverage to thousands of markets while maintaining the same level of rigor.

Currently, most agents are still at the analyst to copilot level, as many are still in the testing phase.

Yield Optimization: Impressive Agent Performance

Liquidity provision has been a realm where automation has been prevalent, with the total TVL held by agents exceeding $39 million. This figure primarily measures assets deposited directly by users into agents but does not include capital routed through treasuries.

Giza Tech is one of the largest protocols in this field and launched the first agent app ARMA at the end of last year, aimed at enhancing yield capture for major DeFi protocols. It has attracted over $19 million in managed assets and generated over $40 billion in agent trading volume.

The high ratio of trading volume to managed assets indicates that agents frequently rebalance capital, allowing for higher yield capture. Once capital is deposited into the contract, the execution is automated, providing users with a simple one-click experience requiring minimal supervision.

The performance of ARMA is quantifiably excellent, generating over a 9.75% annualized yield for USDC. Even considering additional rebalancing fees and a 10% performance fee for the agent, the yield surpasses the usual borrowing rates on Aave or Morpho. However, scalability remains a key concern as these agents have not yet been battle-tested to manage or scale to the size of major DeFi protocols.

Transaction: Human Leads Significantly

However, for more complex actions such as transactions, the outcomes are much more varied. The current transaction model operates based on human-defined inputs and provides outputs according to preset rules. Machine learning has advanced this by enabling the model to update its behavior based on new information without explicit reprogramming, moving it towards a co-pilot role. With fully autonomous agents joining, the transaction landscape is set for a significant shift.

Several competitions have been held between agents as well as between humans and agents, revealing significant differences in performance among models. Trade XYZ hosted a human versus agent transaction competition for stocks listed on its platform. Each account started with $10,000 in initial funds with no restrictions on leverage or trading frequency. The results overwhelmingly favored humans, with the top human performers outperforming the top agents by over 5 times.

Meanwhile, Nof1 conducted a competition between models, where several models (Grok-4, GPT-5, Deepseek, Kimi, Qwen3, Claude, Gemini) competed against each other, testing different risk profiles from capital preservation to maximum leverage. The results revealed several factors that could help explain performance differences:

Position Holding Time: There was a strong correlation, with models holding each position for an average of 2-3 hours performing significantly better than models with frequent turnover.

Expectancy: This measures whether the model is profitable on average per trade. Interestingly, only the top 3 models had a positive expectancy, indicating that most models had more losing trades than winning ones.

Leverage: Models operating at an average lower leverage of 6-8 times proved to perform better than those running at over 10 times leverage, as high levels accelerated losses.

Strategy Hints: The Monk Mode model has been the top performer so far, while the Situational Awareness model has performed the worst. Based on the models' characteristics, it shows that focusing on risk management and fewer external inputs leads to better performance.

Base Model: Grok 4.20 significantly outperformed other models by over 22% across various strategy hints and was the only model to be consistently profitable.

Other factors such as long/short preference, trade size, and trustworthiness score did not have enough data or were shown to have any positive correlation with model performance. Overall, the results indicate that agents often perform better within a clearly defined constraint, implying that human oversight is still very much needed in target allocation.

Assessing an Agent

Given that agents are still in the early stages, there is currently no comprehensive evaluation framework. Historical performance is often used as a benchmark for assessing an agent, but it is influenced by fundamental factors that provide stronger indications of robust agent performance.

Performance under Different Volatilities: This includes disciplined loss control when conditions worsen, indicating that the agent can identify off-chain factors that affect trading profitability.

Transparency vs. Privacy: Both sides have their trade-offs. A transparent agent, if tradable by replication, essentially has no strategic advantage. A private agent faces the risk of creator front-running, where the creator can easily frontrun their own users.

Source of Information: The data sources accessed by the agent are crucial in determining how the agent makes decisions. Ensuring the sources are reliable and free from single dependencies is crucial.

Security: Having smart contract audits and appropriate fund custody architecture to ensure backup measures in a black swan event is crucial.

Next Steps for Agents

For mass adoption of agents, there is still a lot of infrastructure work to be done. This can be boiled down to key issues surrounding agent trust and execution. Autonomous agents' actions are unfenced, and instances of poor fund management have already emerged.

ERC-8004 went live in January 2026, becoming the first on-chain registry that allows autonomous agents to discover each other, establish verifiable reputations, and collaborate securely. This is a key unlocking of DeFi composability, as trust scores are embedded within the smart contract itself, allowing for permissionless interactions between agents and protocols.

However, this does not guarantee that agents will always operate non-maliciously, as security vulnerabilities such as collusion of reputation and sybil attacks could still occur. Therefore, there is still a significant gap to be filled in terms of insurance, security, and the economic staking of agents.

As DeFi agent activity expands, strategy congestion has become a structural risk. Yield farming is the most obvious precedent, where as strategies proliferate, returns compress. A similar dynamic may apply to agent trading. If many agents are training on similar data and optimizing for similar objectives, they will converge on similar positions and similar exit signals.

A version of this issue was formalized in a January 2026 CoinAlg paper from Cornell University. Transparent agents are susceptible to arbitrage as their trades are predictable and front-runnable. Private agents mitigate this risk but introduce a different one, where creators retain an informational advantage over their users and can extract value from the opacity that was originally meant to protect internal knowledge.

Agent activity will only continue to accelerate, and the infrastructure laid today will determine how the next stage of on-chain finance operates. With the increasing utilization of agents, they will self-iterate and become more attuned to adapting to user preferences. Therefore, the primary differentiating factor will come down to trusted infrastructure, which will garner the largest market share.

Original Article Link