Agent tool path evals

Hugging Face argues that agent-facing tools should be evaluated on the path an agent takes, not just whether the final answer is correct. Why it matters: For practical automation, turns, tokens, latency, errors, API calls, and documentation fit can determine whether a tool is dependable enough for recurring agent use.

Try this: When testing one CLI, MCP server, or API with an agent, record the command path, retries, token use, elapsed time, and whether examples led the agent to the intended interface.