How should an agent handle tool failures and retries?

30-second answerSay your answer out loud first, then reveal.

Separate the error types. Transient errors (timeouts, rate limits, 503s) are retried in code with exponential backoff and jitter, without bothering the model. Errors the model can fix (bad arguments, missing fields) are returned to the model with a clear message so it can correct itself. Permanent or unsafe errors (auth denied, resource gone) stop the action and escalate. Never blindly retry non-idempotent actions.

Error taxonomy

Type	Example	Who handles it	How
Transient	429, timeout, 503	Code	Retry w/ exponential backoff + jitter, cap attempts
Model-fixable	Invalid date format, unknown enum value	Model	Return descriptive error and let it retry
Business rule	"Refund exceeds policy limit"	Model, then human	Explain the rule; maybe escalate
Permanent / auth	403, account deleted	Stop	Report to user / escalate, don't loop
Unknown side effect	Timeout on a payment call	Careful!	Check status before retrying (idempotency, Q32)

Good practices

Circuit breakers: if a dependency is down, stop calling it for a while and tell the agent the tool is unavailable so it can plan around it.
Fallback tools: e.g. a secondary search provider.
Error budgets in the loop: after N tool errors in a run, escalate.
Log everything: error type, args and retries in the trace for later analysis.

Common mistakes

Sending a full stack trace to the model. It wastes tokens and can leak internals.
Retrying a "send email" or "charge card" after a timeout and causing a duplicate.

PreviousYour agent has 40+ tools and often picks the wrong one. What do you do?

Every expert started right here.

How should an agent handle tool failures and retries?

Error taxonomy

Good practices

Common mistakes

Related