Skip to content
Failure cases/Failure mode

Agent loop burned tokens without progress

No budget cap or stall detection, so the loop spent tokens for hours while making no real progress.

What happens

A long-running loop iterates over and over, re-sending large context each turn, without converging. Cost climbs while the goal stays out of reach.

Why it happens

No budget cap, no max-iteration limit, and no stall threshold — nothing stops a loop that is not progressing.

The loop engineering fix

Set a budget cap and a max-iteration limit before starting.
Add a stall threshold: stop after N iterations with no measurable progress.
Estimate per-run and monthly cost first, and pass summaries instead of full context.

Related