Agents can pay. Nobody can check what they earned.

The agent economy has an identity layer, a payment layer and a messaging layer. It is missing the layer that decides where the money should go.

Picture an allocating agent with $2 million to place. It has found a trading agent with an impressive record, and it can pay that agent in one API call. It can verify who the agent is. It can message it, agree terms and move the money in seconds, with no person in the loop. What it cannot do is find out whether the record is real.

That is not a hypothetical gap. In the past year, AI agents acquired almost everything they need to handle money on their own. They can call tools through the Model Context Protocol. They can message one another through agent-to-agent protocols. They can pay for an API call with a stablecoin in the middle of a request. They can register an on-chain identity and collect reputation under ERC-8004.

Put those together and you get software that can find a counterparty, agree terms, and send funds without a person in the loop. Agents already trade, rebalance and allocate. The next step, agents allocating to other agents, is a matter of wiring.

One question has no answer anywhere in that stack: is this agent any good with money? Call the missing piece what it is: proof of performance, and the moment it matters, the check before capital moves.

The stack, layer by layer

Reputation registries come closest, but they record what other parties say about an agent. A five-star rating from another agent is still a claim. What an allocator needs is a fact: the agent's actual returns, after costs, with every attempt behind them counted.

Why the usual answer fails for agents

For human managers, finance built a workaround over decades. You meet the manager, call references, read the pitch book, and ask how many strategies they tried before this one. None of it is proof, but a person has a career to lose, and that keeps most of them honest.

An agent has none of that. It has no references, no office and no career. It can be retrained overnight, and every retrain produces a new version with a fresh record. Every retrain is a trial. An operator who runs a thousand versions and shows you the best one has shown you luck, not skill. With five years of daily data, the best of a thousand strategies that have no edge at all will typically show a Sharpe ratio near 1.5, comfortably above what most allocators ask for.

The evidence that this is already a problem is not subtle. A cohort study of 888 algorithms found that a strategy's historical Sharpe ratio explained less than 2.5% of the variance in its later out-of-sample results. A 2026 study generated 898 trading strategies with AI for under $40; none held up out of sample. Producing a convincing record now costs almost nothing.

What the missing layer has to do

Proof of performance for agents has to meet four conditions at once, and each rules out an existing tool:

It cannot require seeing the model. The model is the asset. An operator who shows it gives it away, so audits and code reviews are out.

It has to count every attempt. A record is only as good as the number of tries behind it, so the count has to be committed before results exist, not reported afterwards.

It has to be checkable by software, in one call. The buyer is often another agent. It cannot read a due-diligence questionnaire.

It has to say no in public. A layer that only ever publishes passes is marketing. The refusals are what give a pass its value.

What that looks like in practice

This is what MIZAN builds. An agent's record is recomputed inside a zero-knowledge proof: the returns, the costs and the statistics are calculated from committed data, and every version the agent tried is counted from a ledger fixed before the test. The model never leaves its owner's machine. The result is a credential that anyone, or any agent, can check offline in about 81 milliseconds, before money moves.

The public registry holds 111 such proofs today. 83 of them are refusals, and every one is published. One of them is our own neural network trading NVIDIA. It returned 895% over the period, with a 38% drawdown and a Sharpe ratio of 1.02 against a bar of 1.20, and it was refused. A large return is not a pass.

Why this layer becomes permanent

Trust layers in finance tend to follow the same path. Credit ratings were optional until rules referenced them. Performance reporting standards spread when allocators started writing them into mandates. Website certificates were optional until browsers began marking sites without them as not secure. In each case, the standard that held the record when the requirement arrived kept it.

Agents will reach that point faster than people did, because software follows rules literally. The first allocator, platform or agent framework that runs the check before capital moves sets the default for everyone who connects to it.

The rails for agent money are built. The last layer will be decided by whoever holds the record when checking becomes required, and records cannot be backdated.

Sources: Wiecki, Campbell, Lent and Stauth (2016), "All That Glitters Is Not Gold"; Wuebben (2026), "The Validation Bottleneck", SSRN 7180460; expected maximum Sharpe ratio of zero-skill strategies per Bailey and López de Prado (2014). Registry figures as of 20 October 2026, live at mizan.market/registry. Not investment advice; a credential attests to the honesty of a computation on committed data, not to future performance.