Agent evals need trace and cost checks

Anthropic's agent-evaluation guidance treats quality as the combined behavior of the model, harness, tools, and environment rather than a model score alone. Why it matters: Reliable agents need repeatable checks for task success, tool errors, latency, token usage, and cost before they are trusted with recurring work.

Try this: Create a three-case eval set for one agent task and grade both the final artifact and trace-level signals such as missing sources, failed tool calls, and excessive token use.