Tip: Anthropic Harness Pattern — Separate Evaluator Agent Eliminates Self-Praise, Use Playwright MCP for Grading

Anthropic Engineering Blog·high signal

Anthropic's harness blog reveals that generator agents 'confidently praise their work — even when quality is obviously mediocre.' The fix: a separate Evaluator agent using Playwright MCP to interact with running applications, grading against four explicit criteria (design quality, originality, craft, functionality). Key subtlety: evaluation criteria wording steers generation — phrases like 'museum quality' push toward visual convergence, and 'explicitly penalize purple gradients over white cards' is needed to avoid AI aesthetic defaults.