DWF Deep Report: AI in DeFi Outperforms Humans in Yield Optimization, but Complex Trades Still Lag Behind by 5 Times

Original Title: Will Agents take over DeFi?

Original Source: DWF Ventures

Original Translation: Deep Tide TechFlow

Key Points

Automation and agent activities currently account for about 19% of all on-chain activities, but true end-to-end autonomy has yet to be achieved.

In narrow, well-defined use cases like yield optimization, agents have shown performance superior to humans and bots. However, for multifaceted actions 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 regarding trust and execution, including witch attacks, strategy crowding, and privacy trade-offs.

Continuous Growth of Agent Activities

Over the past year, agent activities have steadily increased, with both trading volume and number of trades on the rise. We have seen Coinbase's x402 protocol leading significant developments, with players like Visa, Stripe, and Google also joining in to launch 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 trading has received 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" endgame, where agents can self-finance, self-execute, and continuously optimize based on changing conditions, has not yet been realized.

Agents are not entirely new to DeFi. For years, there has been automation through bots in on-chain protocols, capturing MEV or achieving excess returns that cannot be realized without code. These systems operate very well under clearly defined parameters that do not change frequently or require additional oversight.

However, the market has become more complex over time. This is where we see the next generation of agents entering, with on-chain activities becoming a testing ground for such developments in recent months.

Actual Performance of Agents

According to reports, agent activities have grown exponentially, with over 17,000 agents launched since 2025. The total amount of automation/agent activity is estimated to cover more than 19% of all on-chain activities. This is not surprising, as it is estimated that over 76% of stablecoin transfer volume is generated by bots. This indicates that there is significant growth potential for agent activities in DeFi.

Agent autonomy exists on a broad spectrum, from chatbot-like experiences that require high human oversight to agents that can formulate strategies based on goal inputs and adapt to market conditions. Compared to bots, agents have several key advantages, including the ability to respond to and execute new information in milliseconds, as well as the capability to expand coverage to thousands of markets while maintaining the same rigor.

Currently, most agents are still at the analyst to co-pilot level, as most are still in the testing phase.

Yield Optimization: Agents Perform Exceptionally

Liquidity provision is a field where automation has frequently occurred, with a total TVL held by agents exceeding $39 million. This figure primarily measures the assets directly deposited by users into agents, excluding capital routed through vaults.

Giza Tech is one of the largest protocols in this field, having launched the first agent application 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 $4 billion in agent trading volume.

The high ratio of trading volume to total managed assets indicates that agents frequently rebalance capital, enabling higher yield capture. Once capital is deposited into the contract, execution is automated, providing users with a simple one-click experience that requires almost no oversight.

The performance of ARMA is measurably exceptional, generating over 9.75% annualized yield for USDC. Even considering additional rebalancing fees and the agent's 10% performance fee, the yield still exceeds that of ordinary lending on Aave or Morpho. Nevertheless, scalability remains a key issue, as these agents have yet to be battle-tested to manage or scale to the size of major DeFi protocols.

Trading: Humans Significantly Ahead

However, for more complex actions like trading, the results are much more varied. Current trading models operate based on human-defined inputs and provide outputs according to preset rules. Machine learning expands this by enabling models to update their behavior based on new information without explicit reprogramming, advancing them to a co-pilot role. The trading landscape will undergo significant changes with the introduction of fully autonomous agents.

Several trading competitions have been held between agents and between humans and agents, revealing substantial differences between models. Trade XYZ hosted a trading competition on its platform for stocks, pitting humans against agents. Each account had an initial capital of $10,000, with no restrictions on leverage or trading frequency. The results overwhelmingly favored humans, with top human performance exceeding that of top agents by more than five times.

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

Holding Time: There is a strong correlation, with models that hold positions for an average of 2-3 hours significantly outperforming those that frequently flip positions.

Expected Value: This measures whether models make money on average per trade. Interestingly, only the top three models had a positive expected value, indicating that most models had more losing trades than winning ones.

Leverage: A lower average leverage level of 6-8 times proved to perform better than models running over 10 times leverage, as high levels accelerate losses.

Prompt Strategies: Monk Mode is the best-performing model so far, while Situational Awareness performed the worst. Based on the characteristics of the models, it shows that focusing on risk management and fewer external sources leads to better performance.

Base Models: Grok 4.20 significantly outperformed other models by over 22% across different prompt strategies and was the only model to average a profit.

Other factors such as long/short preferences, trade sizes, and confidence scores did not have enough data or were proven to have any positive correlation with model performance. Overall, the results indicate that agents tend to perform better within clearly defined constraints, suggesting that humans are still very much needed in goal configuration.

How to Evaluate Agents

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 evaluating agents, but it is influenced by underlying factors that provide stronger signals of robust agent performance.

Performance under Different Volatility: This includes disciplined loss control when conditions deteriorate, indicating that agents can identify off-chain factors that affect trading profitability.

Transparency vs. Privacy: Both sides have their trade-offs. Transparent agents, if they can be actively copied in trades, essentially have no strategic advantage. Private agents face the risk of internal extraction by creators, who can easily front-run their own users.

Information Sources: The data sources accessed by agents are crucial for determining how agents make decisions. Ensuring sources are reliable and not singularly dependent is vital.

Security: It is essential to have smart contract audits and appropriate fund custody structures to ensure backup measures in the event of black swan events.

Next Steps for Agents

To achieve large-scale adoption of agents, there is still a lot of work to be done in terms of infrastructure. This boils down to key issues surrounding trust and execution of agents. Autonomous agents operate without guardrails, and instances of poor fund management have already emerged.

ERC-8004 is set to launch 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 unlock for DeFi composability, as trust scores are embedded within the smart contracts themselves, allowing for permissionless activities between agents and protocols.

This does not guarantee that agents will always operate in a non-malicious manner, as vulnerabilities such as collusion reputations and witch attacks may still occur. Therefore, there remains significant room to fill in areas such as insurance, security, and economic staking of agents.

As agent activities expand in DeFi, strategy crowding becomes a structural risk. Yield farming is the most explicit precedent, as returns compress with the proliferation of strategies. The same dynamics may apply to agent trading. If a large number of agents train on similar data and optimize for similar goals, they will converge on similar positions and similar exit signals.

The CoinAlg paper published by Cornell University in January 2026 formalizes one version of this issue. Transparent agents can be arbitraged because their trades are predictable and can be front-run. Private agents avoid this risk but introduce a different risk, where creators retain informational advantages over their own users and can extract value through opacity from internal knowledge that was meant to be protected.

Agent activities will only continue to accelerate, and the infrastructure laid today will determine how on-chain finance operates in the next phase. As the usage of agents increases, they will self-iterate and become more adept at adapting to user preferences. Therefore, the primary differentiating factor will boil down to trustworthy infrastructure, which will capture the largest market share.