Skills
Anthropic Three-Agent Harness: Planner-Generator-Evaluator Architecture for Multi-Hour Autonomous Builds
Anthropic published a new engineering blog (April 4) detailing a three-agent harness—Planner, Generator, Evaluator—for long-running autonomous app development. The Planner expands prompts into detailed specs, the Generator builds in sprints, and the Evaluator uses Playwright MCP to click through the live app and score against four criteria (design, originality, craft, functionality). A solo run costs $9/20min; the full harness costs $200/6hrs but delivers production-grade full-stack apps. Key insight: the Planner must be deliberately ambitious—conservative planning consistently produces underwhelming results.
↳ Follow the thread