Tool Execution Design

Agents retry. They retry on timeouts, on ambiguous errors, and sometimes on nothing at all. Execution design is about making that safe: bounded time, safe repetition, and consistent state when a multi-step operation fails halfway.

Procedure

1. Measure or estimate the wall-clock time of the operation at p50 and p99.
2. Pick the access pattern: synchronous for bounded seconds, async job for minutes, streaming when partial output is useful as it arrives, event-driven when the tool notifies rather than answers.
3. For every command tool, decide the retry story: natural idempotency, an idempotency key, or a documented "not safe to retry" with a confirmation gate.
4. For every multi-step command, decide the consistency story: a real transaction where one system owns all steps, or compensation where several systems do.
5. Set the timeout and decide what a timeout returns.

Rules with examples