AI Capital Markets
An autonomous agent with a wallet is not a novelty. It is a market participant with no legal personality, no reputation and no counterparty who can call it.
The problem
Software has been allocating capital for decades, but always inside an institution that carried the legal and reputational weight. An autonomous onchain agent has neither. It can hold assets, sign transactions and take positions, and there is no established answer to who is accountable when it does something ruinous.
The market’s current answer is a multisig and a hope. That does not scale to agents operating continuously across venues at machine speed.
Mandates
The core primitive is the mandate: an onchain, machine-readable statement of what an agent is permitted to do with whose capital, enforced at the settlement layer rather than trusted to the agent’s own code. Venue whitelists, position limits, drawdown ceilings and time bounds are constraints the chain checks, not instructions the agent is asked to follow.
An agent that tries to exceed its mandate does not misbehave; its transaction simply does not settle.
Attribution
If several agents operate against a shared pool, performance has to be attributable to each of them, or the pool cannot price them. We record per-mandate position and P&L attribution as chain state, which makes an agent’s track record a verifiable object rather than a claim on a dashboard.
Where this goes
The interesting question is not whether agents can trade — they can — but whether a market can form around allocating capital to them on the strength of verifiable history. That is what this infrastructure is for, and it is the subject of the lab’s ongoing work on agents as capital allocators.