The runaway loop is the real bill
An agent that retries a failing tool call ten times has just spent ten times your intended budget, often overnight. The fix is not to monitor harder — it is to put a hard limit in front of the loop.
Three levers that work
- Token budgets — a ceiling per task or per day.
- Step budgets — a maximum number of plan-act-reflect cycles.
- Spend caps — a dollar amount the agent cannot exceed.
Where to see it
The dashboard shows spend per agent, per day, and per model. Blocked prompt-injection attempts are counted separately, because those are calls you never paid for.
Set it up in a minute
- Open Settings and set a default daily cap.
- Give each agent a per-run token budget.
- Turn on the loop limit so a stuck task stops itself.
- Check the spend tile at the end of the day.
Cheapest key that fits
The platform can route each task to the cheapest provider key that still fits the job. A quick file listing does not need your most expensive model — and now it will not use it.