Wink Pings

The model is powerful, but agents still fail in real repositories

An open-source course uses 14 lectures and 8 projects to break down the coding agent harness: instructions, state, verification, scope, session lifecycle, and how to choose between loops and graphs.

Sumanth shared an open-source course on X, learn-harness-engineering. It has 14 lectures and 8 projects, and the core question is very concrete: how do you get a coding agent to work continuously in a real repository for several hours and then pick up from where it left off in the next session?

Sumanth 分享的课程截图

The course maintainers define harness as the system outside the model. The model decides what code to write; the harness decides when it writes, where it writes, how the work gets verified after it is written, and where the next session starts. There is a blunt line in the README: the strongest model in the world will still fail real engineering tasks without the right environment.

They cite Anthropic's controlled experiment. Same Opus 4.5, same prompt. Without a harness, it cost $9 and took 20 minutes, and the artifact would not run; with a complete harness made up of planner, generator, and evaluator, it cost $200 and took 6 hours, and the resulting game was playable. The model did not change; the environment did. OpenAI's report on Codex is similar: in a repository with a well-configured harness, the same model goes from unreliable to reliable.

harness 模式

This system is broken into five subsystems: instructions, state, verification, scope, and session lifecycle. Instructions tell the agent what to do in what order and which files to read before starting. State writes progress to disk so the next session can resume. Verification only trusts tests, lint, type checks, and the full pipeline. Scope limits the agent to one feature at a time, preventing it from spreading itself across three features at once. Session lifecycle handles initialization before starting, cleanup at the end, and leaves a clean recovery path.

五个子系统

The course projects are progressive. P01 compares a plain prompt with an agent that has repository rules. P02 restructures the repository into something the agent can read. P03 adds multi-session continuity. P04 adds runtime feedback and scope control. P05 has the agent verify its own work. P06 assembles the complete harness. P07 builds automated loops, including goal loops, timer loops, and maker-checker. P08 draws the workflow as a graph. Each project's solution is the starter for the next project, and the same Electron knowledge-base app evolves throughout.

More interesting is how the course handles loops and graphs. In another article, Sumanth mentions that in July 2026, Peter Steinberger asked on X whether we were still talking about loops or had already moved on to graphs. Hamel Husain then wrote Loop Engineering Is Dead. Enter Graph Engineering. Both were poking fun at the industry for renaming the same set of concepts. Underneath the joke is a real choice.

Sumanth's judgment: a loop is a node inside a graph. The two are not opposed. A loop fits a single coherent task: one agent, a clear goal, verifiable output, and a stopping condition. Use a graph when the work splits into parallel tracks, when different agents need different context, memory, or permissions, or when you need auditable control flow. Run a loop first; run it until it breaks, and let the failure tell you which node to split off next. One easily overlooked point: if the loop itself is broken, switching to a graph will not fix it. You will just get a more complex system with the same problems.

The industry's habit of renaming things is also worth watching. Prompt engineering became context engineering, then harness engineering, loop engineering, and graph engineering. Sumanth says that most of the time it is the same work described from a different distance. LangGraph was released in January 2024, and Microsoft AutoGen and Google ADK followed. Multi-agent systems, explicit state, and control flow have long been documented. The terminology is new; the engineering practice is not.

The course also gives you things you can take away directly: AGENTS.md, init.sh, feature_list.json, progress logs, maker-checker prompts, and the harness-creator skill. The repository includes an audit-harness.sh with zero dependencies that checks whether an existing repository conforms to the five subsystems. It is at github.com/walkinglabs/learn-harness-engineering, MIT licensed, in 15 languages.

Who it's for: people already using Claude Code, Codex, or another coding agent who want it to be more reliable in real projects. Not for people looking for a zero-code introduction to AI, and not for people who only care about prompts. The course assumes you can have the agent modify files, run commands, and inspect output. If you can't do those things, reading it will only give you a surface-level view.

The last word goes to verification. In the course, a reader suggested a very practical test: stop the agent halfway through and see whether the next session can continue without starting over. If it can pick up, the harness is real.

发布时间: 2026-11-19 08:32