A watchdog every 60 seconds
A loop checks the desk once a minute: is it alive, is the kill switch set, are the positions protected. If the service crashes, the system restarts it 5 seconds later.
Work · Case study
We run our own trading desk with an AI agent in charge, around the clock, with nobody watching the screen. This page is about the system that keeps an agent like that safe: what it is allowed to do, what stops it, and how every "no" is recorded.
This page describes software controls only. It reports no results, it is not investment advice, and we do not offer trading signals.
A loop checks the desk once a minute: is it alive, is the kill switch set, are the positions protected. If the service crashes, the system restarts it 5 seconds later.
The agent proposes; the guardrails decide. Each trade must pass a daily loss limit, a cap on trades per session, a maximum number of open positions, a cooldown between trades, a per-trade loss limit and a watch list.
The money at risk across all open positions, plus the new one, must stay within a fixed share of the account. If it would not, the trade is skipped.
Three steps, measured from the peak: full risk, then half risk, then no new entries at all. A deeper limit closes everything and stops the desk.
A stop file, or a one-word stop from a human, halts the desk. Messages from other agents cannot trip it or clear it. Restarting takes a deliberate re-arm by a person.
When a guardrail says no, the reason is written down. In one archived log (26 July to 21 September 2026) the desk refused 1,994 trades: 1,823 because the desk was full, 74 at the daily give-back limit, 57 for symbols off the watch list, 40 at the daily loss limit.
When something goes wrong, the lesson is turned into a rule and recorded in a ledger: what happened, which guardrail now prevents it. That ledger has 16 rows today.
Trading is just the harshest test: real money, every minute, no undo. The pattern carries over to any job you want an agent to do on its own: operations, finance back office, customer support. We build the agent and the guardrails around it: limits it cannot argue with, a watchdog, a kill switch a person controls, and a log of every refusal with its reason.
Email us the workflow → · Run the desk yourself · All services