AI Agents as Capital Allocators
Autonomous agents can already hold assets and take positions. What does not yet exist is the market structure that would let capital be allocated to them on terms anyone could defend. We argue the missing primitives are enforced mandates and verifiable attribution, describe an implementation of both, and examine the reflexivity problem that arises when many agents share a training lineage.
The accountability gap
Software has allocated capital for decades inside institutions that carried the legal and reputational weight of its decisions. An autonomous onchain agent has neither a legal personality nor a counterparty who can be called. The prevailing answer — a multisig and a well-intentioned operator — does not scale to agents operating continuously across venues at machine speed.
Mandates as enforced constraints
We describe the mandate as a machine-readable, onchain statement of what an agent may do with whose capital, enforced at settlement rather than trusted to the agent’s own execution. Venue whitelists, position limits, drawdown ceilings and time bounds become conditions the chain checks.
The distinction matters more than it may appear. An agent asked to respect a limit can fail to; an agent whose out-of-mandate transaction does not settle cannot. This also removes the need to reason about the agent’s internals, which is fortunate, because for most interesting agents nobody can.
Attribution and track record
Capital cannot be priced to an agent without a verifiable history. Per-mandate position and P&L attribution recorded as chain state turns an agent’s record into an object that can be independently reconstructed rather than a claim on a dashboard. We treat this as the precondition for anything resembling an allocation market.
Reflexivity
The result we find most uncomfortable: in simulation, agent populations sharing a training lineage exhibited correlated positioning that neither their individual mandates nor their published track records revealed. Diversifying across agents produced far less diversification than the allocator believed they had bought.
We do not have a good answer. Mandate-level exposure limits help and do not solve it. We think this is the most important open problem in the area and we would rather state it plainly than gesture past it.