Open-source Harness Engineering course hits 19k stars, shows $9 vs $200 gap between failing and working agent setups
thisguyknowsai · x · 2026-10-15
A fully open-source, project-based course called Learn Harness Engineering (19k+ GitHub stars, MIT license) is going viral for teaching how to build the environment around AI coding agents. Its core claim, backed by an Anthropic experiment with the same model (Opus 4.5) and prompt: without a harness, $9 and 20 minutes produced something that didn't work; with a harness, $200 and 6 hours produced a playable game. Same model, different harness.
The course breaks a harness into five subsystems — instructions (AGENTS.md/CLAUDE.md), tools, environment, state, and feedback — across 14 lectures and 8 projects, translated into 15 languages including Simplified Chinese.
A new "Frontier Harness Design Breakdowns" section reverse-engineers how four frontier products build real harnesses: Claude Code (four-layer memory, five-level compaction, hooks, sub-agent isolation), Codex (repo as source of truth, worktree isolation), Pi (minimal kernel, programmable expansion), and DeepSeek ("everything is a plugin," event pipeline).
More from coding & agent
- Bun ships bun lint: 26x faster ESLint-compatible linter; Oxlint creator unimpressed — cnakazawa · 2026-10-15
- REA 'Reverse Engineer Anything' tops GitHub trending, letting Claude Code and Codex decompile any app — blaizedsouza · 2026-10-15
- REA v6.5 ships with 529 PRs from 200+ contributors, adding offline evidence inspection and JS bundle recovery — blaizedsouza · 2026-10-15
- Cline: Gemini 4 Argon saw 3x Claude's day-one adoption and 30% fewer tokens — JMateosGarcia · 2026-10-15
- Codex Mac app hit 5M weekly users in 4 months, then became the ChatGPT app — every · 2026-10-15
- AI coding agents on production codebases: 4 setups tried, 3 failed, one workflow stuck — paulabartabajo_ · 2026-10-15