Project-based course on building harnesses for Codex and Claude Code, covering environment design, state management, verification, and control systems.
Key Takeaways
Harnesses don’t improve model intelligence; they create closed-loop systems with explicit rules, boundaries, and observable runtime behavior.
Core curriculum covers stopping agents from declaring victory early, maintaining context across multi-session tasks, and verifying work with full-pipeline tests.
References primary sources from OpenAI and Anthropic on harness engineering for long-running agents and application development.
Starter templates include AGENTS.md, feature_list.json, and claude-progress.md for immediate use in your own projects.
Structured into lectures, hands-on projects, and a copy-ready resource library, with Project 01 walking through baseline vs. minimal harness comparison.
