Retry with backoff: make flaky APIs agent-safe

<p>APIs fail for a moment and then recover. An agent should wait and try again, and it should not create a second record while it does.</p><h2>Worked example</h2><p>An agent writes notes to a CRM that sometimes returns 429 or 500.</p><ol><li>Retry only 408, 429, and 500–504. Stop on 400, 401, 403, and 404.</li><li>Wait longer each time, with a little randomness so every agent does not retry together. Start around 500ms, double it, and stop at 30 seconds.</li><li>Cap it at four tries, then stop calling that endpoint for a while.</li><li>Send an idempotency token so a repeated write does not create a duplicate.</li><li>Keep the failed job, with the original input, so someone can replay it.</li></ol><h2>Common mistakes</h2><ul><li>A fixed delay, so every retry hits at the same instant.</li><li>Unlimited retries that run up the bill.</li><li>Retrying a write that is not safe to repeat.</li><li>No randomness in the wait.</li><li>No log of which calls failed.</li></ul><h2>Try it</h2><p>Open the <a href="https://gowithagentic.ai/tools/retry-backoff">retry and backoff</a> tool.</p>