Anthropic Harness Design for Long-Running Apps: Three-Agent Architecture (Planner + Generator + Evaluator) With Progress Files Spanning Multi-Hour Sessions
Anthropic Engineering·medium signal
Anthropic's engineering blog details a three-agent harness for tasks spanning hours or days: separate planning, generation, and evaluation agents with a claude-progress.txt file for cross-session state. Key insight: agents overrate their own output on subjective tasks — a separate evaluator agent calibrated with few-shot examples prevents quality drift. First context window gets a different initializer prompt to establish environment state.