Agent-as-evaluator approaches use an agent (or multi-agent system) to judge outputs and processes of other agents.
Why It Matters
Some evaluation tasks themselves require tools and multi-step reasoning (running code, checking docs, reproducing bugs).