Skills
Anthropic Publishes Three-Agent Harness: Planner/Generator/Evaluator with GAN-Inspired Separation for Multi-Hour Coding
Anthropic's engineering blog details a three-agent harness separating planning, generation, and evaluation for multi-hour autonomous coding sessions. The Planner decomposes specs into self-contained tasks, the Generator builds code within a single context window per task using structured handoff artifacts (JSON specs, claude-progress.txt), and the Evaluator grades output using Playwright MCP to navigate live pages. Inspired by GANs, the key insight is that separating the agent doing work from the agent judging it is the strongest lever for quality.
↳ Follow the thread