Agentic and algorithmic trading evaluations

Most prop firms treat algorithmic trading as a policy question: permitted, restricted, or banned, with rules about copy trading and latency arbitrage layered on top. That framing hides the more interesting problem, which is technical, not administrative.

If an evaluation promises that its verdicts can be reproduced, then letting a machine trade it is a claim about determinism — and determinism is much harder to hold at machine order rates than at human ones.

Why determinism is the binding constraint

A reproducible verdict means the same inputs always reach the same result. Replay the recorded prices and the recorded orders and you land on the identical equity curve, the identical breach, the identical verdict.

Every time, on any machine. That is the promise being made.

That is straightforward when orders arrive seconds apart. It becomes delicate when they arrive in bursts, because anything that makes ordering depend on arrival rather than on recorded data breaks reproducibility in a way that is almost invisible: the replay produces a slightly different answer, and only sometimes.

What an evaluation must guarantee before allowing machine traders

1. A total ordering over events derived entirely from recorded data, with no residual dependence on arrival order.
2. Idempotent order submission, so a retried request cannot become a second position.
3. Replay cost that stays tractable as order count grows, or a verdict cannot be finalised at all.
4. Attribution that distinguishes a machine-submitted order from a human one, in the audit record.
5. Risk gates that hold on every path, since an agent will find an unguarded one far faster than a person.

The fifth point deserves emphasis. Human traders explore a rulebook slowly and mostly by accident. An automated strategy explores it systematically, at rate, and will reach an unguarded edge case in hours. Any gate that exists only on the interface a person uses is not a gate.

Why the API surface is the easy half

Publishing endpoints for placing and cancelling orders is ordinary engineering. The genuinely hard part is everything behind them: guaranteeing that an evaluation traded at machine rate still produces a verdict anyone can reproduce, and that the replay proving it finishes within the resources available.

A firm that cannot reproduce a machine-traded verdict has quietly stopped offering the thing it advertised. It would still issue verdicts — they simply could not be checked, which returns the trader to trusting the firm's word and undoes the entire point.

What this means if you trade algorithmically

• Ask whether verdicts on machine-traded accounts are reproducible, not merely whether bots are permitted.
• Ask how simultaneous events are ordered. "By timestamp" is an incomplete answer — timestamps tie.
• Ask whether order submission is idempotent, and what happens to a retry after a network timeout.
• Prefer a firm that documents these constraints over one that simply says algorithmic trading is allowed.

The interface for programmatic access on this site is documented publicly, and the determinism constraints above are the reason it is specified carefully rather than opened broadly. An evaluation is worth only as much as the reproducibility behind it.