Testing AI Agents: Production IS Your Test

Traditional test suites fall short for AI agents because outputs are non-deterministic and user interactions are unpredictable. The recommended approach is treating production as the primary evaluation environment: collecting observability traces from real interactions (user queries, tool calls, responses, retries) and periodically analyzing them to identify failure patterns such as malformed tool calls, hallucinations, and weak request categories. These insights then drive improvements to system prompts, tool descriptions, model selection, and knowledge bases.