Harness Evals Framework for AI Agents

Agent evaluation is consolidating around golden workloads, trajectory and conversation metrics, tool accuracy, deterministic boundaries, refusal and escalation paths, handoff fault injection, retrieval-versus-generation attribution, durable-state recovery, replay, observability, and cost-aware CI gates. FlowAgent adds a production CI repair example with generate-and-validate loops, abstention, adoption metrics, and human correctness review; broader reproducibility and transfer remain open.