Whenever you're ready, walk me through how you'd approach this.
You are evaluating an AI agent that retrieves context, uses tools, and produces multi-step answers. Exact string matching is clearly too narrow because the agent can succeed with different wording, partially fail in retrieval, or take an inefficient tool path while still sounding correct.
How do you evaluate the performance of an AI agent beyond simple output matching (e.g., using RAGAs or custom evaluation frameworks)?