Within Anthropic's three-agent harness, the Evaluator agent is calibrated with few-shot examples and explicit scoring criteria to handle subjective assessments (does this UI look good? is the UX intuitive?) while maintaining reproducibility for objective tasks. It navigates live pages using Playwright MCP, interacts with the interface, and provides detailed critiques that guide the Generator in iterative improvement cycles. This separates the 'build' from the 'judge' to avoid the self-evaluation trap where generators rate their own output too highly.