How We Are Testing Our Agents in Dev

Testing AI agents is difficult due to their non-deterministic and unstructured outputs, making traditional testing methods ineffective. A practical approach involves evaluating agents on dimensions like semantic distance, groundedness, and tool usage, often leveraging an LLM-as-judge. To handle the variability, a concept of "soft failures" is used, where scores within a certain range are tolerated unless a specific threshold is exceeded, which then constitutes a hard failure. Other best practices include re-evaluating soft failures, requiring LLM judges to explain their scores for better debugging, and removing unreliable or "flaky" tests. While testing in development is challenging, monitoring agent performance in a live production environment presents an even greater set of problems.