Can we replay what an AI trading agent actually did?

On June 2, Robinhood opened a door that had stayed shut since retail brokerages began: it let an outside AI agent place trades inside a real, funded account. We connect an agent we already use, Claude or ChatGPT or something homemade, fund a ring-fenced account it can touch and nothing else, and let it work. Ask it to check a portfolio for concentration risk and rebalance, and it does. I read the announcement twice, then asked the question I always end up asking about anything that moves real money. If this agent sells a position on a Tuesday, can I sit down on Wednesday and replay exactly why it did? With a classical trading algorithm, yes. With a language model, no, and the reason hides below the layer anyone shows us.

Start with what “deterministic” actually buys us. A rule-based algorithm is a function. Feed it the same inputs, the order book, the price, the position, and it returns the same order every single time. That property is quietly load-bearing. It is what lets a compliance team replay a flagged trade, what lets us back-test a strategy and believe the result, what lets anyone reconstruct why an order fired.

A language model breaks the function. The obvious culprit is temperature, the dial that controls how much randomness the model samples; turn it to zero and you would expect one fixed answer. It does not hold. Even at temperature zero, the same prompt and the same market snapshot can produce different outputs across runs. The dominant reason, which Thinking Machines Lab pinned down in September 2025, is that the chips running the model batch a request alongside other people’s traffic, and the underlying math is not batch-invariant: the model computes slightly different numbers depending on who else happens to be in the batch. Floating-point addition compounds it, because adding the same numbers in a different order produces a marginally different sum. Usually the wording shifts and nobody notices. Occasionally it tips a borderline call from hold to sell.

This stopped being academic the moment a regulated broker wired these agents to live orders. Robinhood’s own disclosures are candid about the boundary it is drawing. The broker, in its terms, “does not control, supervise, monitor, recommend, or audit the agents customers choose to connect”, and it notes plainly that agents “can misinterpret instructions, act on stale information, and behave unpredictably”. Read that next to the non-determinism and the gap sharpens. Even an investor who wanted to audit the agent could not fully reproduce its decision, because the decision was never reproducible to begin with.

The numbers behind that claim are worth seeing directly rather than taking on faith.

The Financial Industry Regulatory Authority (FINRA, the self-regulator that oversees US brokers) named exactly this in its 2026 oversight report, warning that multi-step agent reasoning can be “difficult to trace or explain, complicating auditability”. The regulator is not worried about the agent being wrong. It is worried about no one being able to show, afterward, how it got there.

None of this makes the tool useless, and alarm is the wrong register. Robinhood’s containment is real: ring-fenced funds, a push notification on every trade, a one-tap disconnect. That bounds how much damage a confused agent can do in an afternoon. What it does not do is make the agent’s reasoning reproducible. Containment is not auditability, and the two get conflated constantly.

So the tool-fit test is concrete. Before connecting an agent to a funded account, we can restore some replayability by pinning the things that drift: a frozen model version rather than “latest”, a fixed seed where the provider exposes one, and a full local log of every prompt and every order kept on our own side. Treat the agent as a discretionary manager whose calls we record, not as a back-tested system we trust to behave the same way twice. The broker keeps the trade blotter. We should keep the reasoning trail, because nobody else is going to.

Revenue at Robinhood grew 15% year over year last quarter, down from 100% two quarters earlier, and opening the platform to agents is partly a search for the next curve. That is a reasonable thing for a business to do. The question I keep turning over is smaller and older than any of it. We spent a century making markets trustworthy by making them reproducible, and the most interesting new tool on the retail stack is the first one we cannot replay. I am not sure we have priced what that quietly costs.