@FakeMaidenMaker: A GitHub open source course called "Learn Harness Engineering" recently made it to the front page of Hacker News (155 points, 5.1k stars). The topic is reliability engineering for AI coding agents—that is, OpenA…

A GitHub open source course called "Learn Harness Engineering" recently made it to the front page of Hacker News (155 points, 5.1k stars). The topic is reliability engineering for AI coding agents—that is, what OpenAI and Anthropic have been talking about for the past few months: harness engineering. https://github.com/walkinglabs/learn-harness-engineering… The author is walkinglabs, an open source team from Tsinghua SIGS. The course consists of 12 lectures + 6 hands-on projects, all revolving around a single capstone project (an Electron knowledge base app). It already has 13 language versions. The core points: 1. Harness Engineering is a new concept but definitely not just marketing: OpenAI officially introduced it in the Codex blog in February 2026, and Anthropic published two companion engineering articles around the same time. This course systematically organizes the scattered engineering content from both companies into a readable Chinese system—it's not inventing something new. 2. The core argument is that the model is not the problem; the environment is: Anthropic ran a controlled experiment: with the same Opus 4.5 and same prompt ("make a 2D retro game editor"), without a harness it cost $9 and ran for 20 minutes with no output; with a full harness (planner + generator + evaluator) it cost $200 and ran for 6 hours to produce a playable game. The difference is not in the model, but in the environment. 3. The 5 subsystems of a harness: Instructions (what the agent sees), State (what has been done), Verification (how to determine completion), Scope (only one feature at a time), Lifecycle (session start and end). 4. Minimum starting point: 4 files: AGENTS.md (operations manual), init.sh (environment check script to run at each start), feature_list.json (machine-readable feature boundaries), claude-progress.md (progress handoff between sessions). Just drop these 4 files into your existing codebase and you're ready to go—no need to finish all 12 lectures. 5. Why agents always mess up (questions the course repeatedly answers): They try to do three things at once but only finish half of each; they say "done" but tests don't pass; there is no memory across sessions—halfway done, then restart from scratch next time; without a verification pipeline there is no true completion (only test pass + lint pass + e2e pass all together counts). 6. 6 projects with progressive evolution: P01: prompt-only vs rules-first comparison to see the gap; P02: restructure the codebase into an agent-readable format; P03: cross-session continuation; P04: add runtime feedback to control scope; P05: make the agent self-verify; P06: full harness with observability. 7. The capstone project is an Electron knowledge base app: The answer to each project becomes the starting code for the next—the app evolves, and your harness skills evolve with it. Summary: The role of "AI agent engineer" is shifting from "someone who can write prompts" to "someone who can build harnesses." This course itself isn't extremely hardcore, but it systematically organizes the scattered engineering content from both OpenAI and Anthropic into a Chinese-indexed system. For the Chinese-speaking community looking to get started in this direction, it saves the time of reading 5 English engineering blog posts.