Preserve Before You Rerun: Critical Steps for Debugging Production AI Agents

Immediately rerunning a failed AI agent masks the actual failure and risks repeating unintended side effects like sending duplicate emails or issuing multiple refunds. This debugging checklist prioritizes capturing evidence first—trace IDs, original context, retrieved documents, and detailed tool call records—before any fix attempt. By systematically preserving this evidence, teams transform production incidents into reproducible regressions that reveal the true root cause.