Looping
When an agent gets stuck repeating the same step, never making progress toward done.
Think of it like
A GPS that keeps rerouting you around the same block, confidently, forever.
Example
The agent tries a broken command, reads the error, tries the exact same command again, reads the same error — and does it forty times until something stops it.
How it actually works
Looping usually happens when the agent can’t tell that its last action failed, or has no new information to act on, so it just re-derives the same plan. It’s the classic agent failure mode and the reason you never run a loop without a ceiling. Fixes range from clearer tool errors to nudging the agent to try something different after a repeat.
For product teams
An agent with no loop ceiling can quietly burn a fortune in tokens doing nothing — always cap it.
For engineers
The agent re-emits equivalent actions because the observation doesn’t change its state enough to alter the next decision.
Related
- Stop Condition — The rule that finally breaks the cycle.
- Max Iterations — The hard cap that ends a runaway loop.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome