Risk controls in autonomous trading

By the TradeAgentic engineering team·Published 30 August 2026·Updated 30 August 2026

The question that actually matters

People evaluating autonomous trading software usually ask how well it performs. The more useful question is what happens when it is wrong, because it will be wrong regularly, and the difference between a bad week and a disaster is entirely in the controls.

A system with no edge and excellent controls loses slowly and survives. A system with an edge and poor controls can lose everything in a single unattended night. Only one of those failures is recoverable.

Where protection belongs

Protection belongs at the broker, not in the software

This is the single most important thing to check, and it is easy to check. Ask whether protective orders are placed at the brokerage, or held in the software's memory and submitted when a level is reached.

If protection lives in the running process, then a crash, a restart, a network drop or a laptop lid closing removes it. The position stays open and unprotected, and nothing tells you. If protection rests at the broker, it survives all of that, because the broker does not care whether your software is running.

The follow up question is what the software does when it cannot place that protection. The only acceptable answer is that it closes the position rather than holding it unprotected.

Limits that end the day

An autonomous system needs a point at which it stops for the day and does not resume. That should be a fixed loss threshold, set by you, checked before every order, and not adjustable by any automated part of the system.

A profit target is worth having too, but only as a stopping rule. If reaching a target causes the software to increase size in pursuit of more, that is not a target, it is an accelerator. Position size should never increase in response to losses either. Any system that does this has a martingale in it, whatever the marketing calls it.

A control set that nothing may weaken

The useful design pattern is a small set of protections that the rest of the system, including any self tuning, is structurally forbidden from altering. Loss limits, the immediate flatten-everything control, broker-resident protection, refusal to trade on failed startup checks, and concentration caps.

Ask specifically whether the system's own optimisation can widen any of these. If it can, they are suggestions.

Proving the controls work

Prove it before it scales

Any system that will eventually trade meaningful size should demonstrate the full path with trivial size first, including that protective orders genuinely arrive at the broker. Early trades running at a fraction of full size, stepping up only once the path has proven itself, costs almost nothing and catches the failures that matter.

The same applies to you. Run it on paper until you have watched it refuse things, recover from a restart, and behave through a day you did not expect.

What no system can protect you from

Gaps. A stop cannot execute at your price if the market opens past it, which is why avoiding known events matters more than any stop placement. Broker outages. Your own risk settings, which are the largest single determinant of outcome and are entirely yours.

And the general case: trading involves risk of loss, and controls bound the damage rather than eliminating it. Any vendor implying otherwise is telling you something untrue.

Data quality is a risk control

Software that cannot see the market properly should not be trading it. A system that keeps operating on stale or degraded data is not being resilient, it is guessing with confidence.

Ask what happens when a feed goes stale. Standing aside is the correct behaviour, and a system that does it will say so on screen rather than quietly continuing.

Risks that are easy to miss

Concentration is the quiet one

Most attention goes to stop placement, but the failure that hurts most is usually too much in one place at once. Caps on how much of the account any single position or asset class may become do more work than any individual stop.

Ask what those caps are and whether they are enforced before every order rather than reviewed afterwards.

What to watch during evaluation

Not the profit and loss. Watch whether it refuses things and whether the refusals make sense. Watch a restart with a position open. Watch a day where the data misbehaves.

A month of paper trading tells you very little about whether a strategy earns and a great deal about whether the software is safe to leave running, which is the question that actually determines your downside.

Putting it together

The controls are not the strategy

It is tempting to evaluate autonomous software on what it trades. The controls are the more important half, and they are far easier to verify, because they either exist or they do not.

A vendor who cannot answer plainly where protective orders live, what stops the day, and what nothing may weaken has told you what you need to know.

Questions worth asking in order

Where do protective orders rest. What happens if one cannot be placed. What halts trading for the day, and can the software widen it. What flattens everything, and how fast. What happens on degraded data. What is checked before the first order of the session.

Six questions. Any vendor selling autonomous trading should be able to answer all of them without preparation.