Agent evaluation guides

Evaluate tool calls, intermediate previews, factual fidelity, text fit, and final delivery separately.