Retries, Timeouts, and Idempotency for Agent Actions
Retries, Timeouts, and Idempotency for Agent Actions
Why agents amplify reliability issues
If a tool call fails, agents tend to “try again”. Without idempotency, that can create duplicate tickets, duplicate payments, or double updates.
Idempotency keys
For any write action, include an idempotency key derived from (user, action, target, timestamp bucket). The server should detect repeats and return the previous result.
Timeouts and circuit breakers
Set timeouts per tool. If a downstream system is failing, open a circuit breaker so the agent stops hammering it.
User-safe behavior
When the system is uncertain, the agent should ask before repeating a write action: “I might have created this already—do you want me to check?”
