Autonomous agents now control nearly one in five on-chain actions — but they still trail humans at open-ended trading, according to new research.
DWF Ventures’ April 16, 2026 report finds that autonomous on-chain agents — software that plans, decides and executes transactions without real-time human input — account for more than 19% of on-chain activity today. These programs already manage over $39 million in value-locked positions and are being used for yield strategies, liquidity management, portfolio rebalances and trade execution. Most deployments, the report notes, remain in early testing phases.
Where agents shine
- Yield optimization has emerged as a clear early win. Agents that move stablecoins and capital between lending protocols to chase the best rates can outperform manual strategies in stable, rule-driven environments. The report highlights Giza’s ARMA agent, which routes stablecoins across lenders and generated about 9.75% annualized returns for users — outpacing yields on protocols such as Aave and Morpho in the sample studied.
- Narrow, well-defined objectives are ideal for agents because the rules and parameters rarely change. “Agents thrive when the objective is narrow and the parameters don’t move often,” said Xin Yi Lim, senior associate for investments at DWF Labs.
Where agents still lose
- In open-ended trading and environments that require context, narrative understanding and real-time adaptation, agents lag behind humans. In a tradexyz stock-trading contest cited by DWF, the top human trader beat the best agent by more than 5x. In another competition run by nof1, only three of seven leading AI models were able to make a profit per trade.
- “Agents struggle when the situation isn’t clearly defined,” Lim told Decrypt. Aytunc Yildizli of 0G Labs added that closing this gap will require more than bigger models: agents need cryptographic proof of action, trusted execution environments and decentralized infrastructure so trust isn’t concentrated in a single cloud provider.
Industry reaction and infrastructure work
- Builders and execs are bullish but cautious. MoonPay chief engineer Neeraj Prasad said an agent “can be as capable as a human if given all the context and tools,” but warned agents will become “more competent, harder working, and malicious in some cases.”
- Ethereum ecosystem work is ongoing to make agents more capable. Earlier this month, decentralized relay Biconomy proposed a standard — backed by the Ethereum Foundation — to let agents batch and run multiple actions across DeFi protocols, easing more complex on-chain workflows.
- Coinbase CEO Brian Armstrong echoed long-term optimism, tweeting that an “agentic economy could be larger than the human economy,” and predicting machine-to-machine payments will boost demand for stablecoins beyond current estimates.
A realistic timeframe
- Despite headline numbers, DWF Labs’ Lim emphasized that most of the 19% figure is narrow bot activity — MEV capture, stablecoin routing and similar tasks — and that truly agentic activity remains a minority. He estimates a five- to seven-year runway before agent-driven volume meaningfully rivals human volume in major financial verticals, with on-chain markets likely to reach that point first because of their permissionless infrastructure.
The takeaway
Autonomous agents are fast becoming a material force in DeFi, proving especially effective at repetitive, rules-based tasks like yield optimization. But when markets demand contextual reasoning and adaptability, humans still have the edge. The next phase will hinge on infrastructure upgrades, cryptographic assurances and safer execution environments that let agents act more autonomously without concentrating trust — or amplifying new vectors for error and abuse.
Read more AI-generated news on: undefined/news
