DeFi as an Agent Playground

About this session

Four builders were asked how much control over capital AI agents should be given in DeFi, and largely talked each other out of the optimistic answer. All four saw agents as a real efficiency gain in monitoring, execution and interface design, and none would hand them unsupervised control of funds. Instead they converged on agents that work inside narrow limits, on data checked before they act, with a person still accountable if something goes wrong.

Ken Ng framed the opportunity around who currently holds the edge. Sophisticated liquidity management, the kind that matters for minimising MEV and loss-versus-rebalancing, is practised by market makers and funds with dedicated teams rather than by individuals adjusting positions by hand, and agents level that field. He was less comfortable on risk, noting that writing a risk ceiling into a system whose purpose is to make money is hard because nobody sets out to earn less.

Gauthier Vila offered the clearest evidence against the optimistic case. His team built an LLM-based system for generating yield, found it underperformed expectations, and replaced it with a rule-based agent, concluding that models were useful for surfacing data and too risky to trust with execution. Lukasz Stoczynski described a boundary drawn deliberately: Mimic has the protocol itself verify on-chain and off-chain data before a task executes, so agents trigger work without owning it. He did not expect risk or vault curation to move to agents soon, for legal rather than technical reasons. Funds under management carry liability, and if an agent lost ten million dollars of someone else's money, who answers for it is unsettled.

Stefano Bury argued the nearest-term win was interface rather than autonomy, pointing to an agentic front end from the Moonwell team that reads like consumer fintech and barely mentions crypto or AI. He reported, on his own unverified figures, that value flowing through multi-agent systems on Virtuals' commerce protocol had reached roughly $130 million within six weeks and was growing about $10 million a day, while conceding orchestration was still human. His longer-term picture had specialised agents competing for each step, with ERC-8004 supplying the trust scores an orchestrator would use to choose between them.