An AI agent retry policy that avoids repeated mistakes

A retry policy defines when an AI workflow may repeat a failed step, how often and when it must stop. Base it on the type of failure and the side effects of the action. A temporary read error and an unknown message delivery should not receive the same treatment.

Classify before repeating

Imagine a document lookup returning a temporary server error. A bounded retry may be reasonable. Now imagine an email send timing out after the request reached the provider. Repeating it without reconciliation could send a duplicate.

Authentication failures, invalid inputs and denied permissions usually need a different action rather than another identical request. A model should not keep trying new routes to bypass an intentional denial.

Write a small decision table

• Temporary read failure: retry within a defined limit and delay policy.
• Invalid input: correct the input or ask for missing information.
• Access denied: stop and route to the authorized owner.
• Unknown write outcome: reconcile before another attempt.
• Repeated failure: preserve the checkpoint and escalate with evidence.

Respect provider rate-limit guidance and avoid synchronized retry bursts. Keep a total task budget so several nested retry loops do not multiply the workload unexpectedly.

Preserve successful steps

If a report was generated but uploading failed, recover the upload rather than regenerate everything. Repeating earlier stages can change the artifact and make it harder to compare with an approval already given.

Keep the original payload and action identity when the retry truly represents the same operation. If the payload changes, treat that as a new proposed action and recheck authorization.

Test the stop condition

Simulate a persistent error and confirm the workflow stops at the agreed limit with a useful handoff. Test an unknown write separately. The desired result is controlled recovery, not an agent that eventually reports success after exhausting every possible path.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.