Risk controls and agent autonomy
The limits that gate every order — kill switch, circuit breaker, caps, autonomy — and how to tune them for how you actually trade.
Перевод пока недоступен — показан английский вариант.
Every order — from the ticket, the CLI or an AI client — passes the same risk gate. These settings live in Settings → Risk limits and are enforced server-side. A client cannot override them, and neither can a CLI flag.
The kill switch #
One toggle that refuses all agent execution immediately. Manual orders from the ticket are unaffected: you can always close what you opened.
Use it when: a strategy misbehaves, a model starts acting strangely, you are stepping away from an automated setup, or you simply want a hard stop while you think.
The circuit breaker #
A second layer that trips after a defined adverse event — consecutive stops or a drawdown threshold — and pauses automated trading until you reset it. The breaker is armed or off per account; when it trips, the app says why and requires an explicit reset.
Caps #
| Cap | What it bounds |
|---|---|
| Max position size | Notional of any single position |
| Max leverage | Multiplier allowed on futures orders |
| Max trades per day | Count of executed orders in a day |
| Max daily loss | Realised loss before execution is refused |
| Max daily notional | Total traded value per day |
Usage is shown next to each cap — trades used, loss used, notional used — so you can see how close you are before an order is refused rather than after.
Autonomy #
Autonomy decides what an agent may do without asking:
| Setting | Behaviour |
|---|---|
| Propose only | The agent can analyse and suggest; nothing executes automatically |
| Ask above threshold | Orders above your max position need your approval; smaller ones may execute |
| Full | The agent may execute within your caps without asking |
Changing autonomy upward requires a confirmation step, because it changes what can happen while you are not looking. Widening it also has a prerequisite: execution tokens cannot be minted while autonomy is Propose only.
Allowlist and blocklist #
Restrict trading to specific pairs, or exclude ones you never want touched. The gate applies to agent execution as well as manual tickets, which makes the blocklist a useful guard against a model mis-reading a symbol.
A configuration that survives contact with reality #
- Set max daily loss to an amount that would make you stop for the day anyway.
- Set max position size to the largest position you would defend at 3am.
- Set max leverage below what the venue allows — margin is the venue's offer, not
your plan.
- Start with Propose only and read what the agent wants to do before letting it act.
- Move to Ask above threshold when the proposals look sane over a week of real markets.
- Keep the kill switch within reach, and test it once so you know what refusing feels like.
What the gate does not do #
It does not pick your trades, size them for you, or stop a losing position from getting worse within your caps. It bounds the damage; the decisions remain yours.
Чек-лист
- Caps set before enabling execution
- Kill switch location known
- Autonomy level matches how closely you watch the account