Check for completion safely
If the tool returns a request or job identifier, query that operation’s status instead of submitting the same command again. Otherwise inspect the affected resource using a read-only call and compare it with the expected state.
For actions such as creating a record or sending a message, use an idempotency key or other stable identity when supported. This lets a retry refer to the same intended operation rather than creating a second one.
Retry with a clear boundary
If evidence shows the action did not complete, retry once the cause is understood. If the state is ambiguous and the action is consequential, pause repeated attempts and gather enough information to avoid compounding the effect.
Record the timeout, operation identity, and final observed state. When a tool cannot expose execution status or support safe retries, treat that limitation as part of the workflow design and add a verification step.