AI Agents Already Run a Fifth of DeFi, But Still Lose to Humans at Trading

In brief

• Autonomous AI agents run 19% of on-chain activity, losing to humans by up to a 5-to-1 margin at open-ended trading, according to a new report.
• Coinbase CEO Brian Armstrong says the agentic economy could soon surpass the human one.
• Researchers argue agents need better infrastructure before they could scale further. New research has found that autonomous agents, software systems that plan, decide, and execute on-chain transactions without direct human input, now drive more than 19% of on-chain activity. But while these agents have outperformed humans on narrow tasks, they still lose by up to a 5-to-1 margin in open-ended trading, according to a DWF Ventures report published Thursday.

— DWF Ventures (@DWFVentures) April 16, 2026 In decentralized finance, these agents run yield strategies across lending protocols, manage liquidity, rebalance portfolios, and execute trades. Total value locked in agent-managed positions has climbed past $39 million, with most deployments still in early testing, per the study. In one sample, autonomous finance protocol Giza's ARMA agent, which moves stablecoins between lending platforms to get the best rates, has earned users 9.75% a year, beating yields on other decentralized finance protocols such as Aave and Morpho, per the study. That picture shifts, however, once the task at hand gets harder. In a stock trading contest run by tradexyz, the top human beat the top agent by more than 5x. A separate contest between leading AI models, held by nof1, found that only three of seven were able to turn a profit per trade. “Agents struggle when the situation isn't clearly defined,” but they thrive “when the objective is narrow and the parameters don't move often,” Xin Yi Lim, senior associate for investments at DWF Labs, told *Decrypt*. This is one of the reasons why yield optimization, the practice of moving capital between lending protocols to capture the highest available returns, has become an early proving ground for agents, Lim explained. "Agents thrive when the objective is narrow and the parameters don't move often, which is why yield optimization works,” Lim said. “Until agents can reason and adapt to real-time information, they will not be able to react when the market changes and conditions are unclear.” Builders in the space appear to echo this concern. An agent can be as capable as a human “if given all the context and tools," MoonPay chief engineer Neeraj Prasad told *Decrypt* in an interview. He warned, however, that "the writing is on the wall that agents are both more competent, harder working, and malicious in some cases."