Claude Code Harness puts the agent on a leash, and that's the point
Everyone who's left a coding agent running unsupervised knows the failure mode. It starts on the task you gave it, then thirty minutes later it's rewritten three files you didn't mention, deleted a test it found inconvenient, and declared victory. The agent didn't get dumber, it just had no leash. Claude Code Harness, trending on GitHub this week, is one answer: stop trusting the agent to behave and build the rails into the workflow instead.
The whole thing is a discipline, not a model. It forces every piece of work through a fixed loop, plan, then work, then review, then release, exposed as five commands: harness-plan, harness-work, harness-review, harness-release, harness-sync. You describe what you want, it drafts a spec.md and a Plans.md with scope, acceptance criteria, and explicit stop conditions, and nothing gets built until you approve that contract. Implementation stays inside the approved boundary. Review runs as a separate step from coding so the thing that wrote the code isn't the thing grading it. Release packages only the verified evidence.
If that sounds bureaucratic, that's the trade. The reason agents drift is that we let them hold the spec, the implementation, and the judgment all in one head. The harness rips those apart and puts a human approval gate between each one. It's the same instinct as code review and CI, just aimed at an agent instead of a junior dev.
It's MIT, it runs on Claude Code 2.1 and up, and it's already deep into the four-point releases, so this isn't a weekend toy. The interesting signal is that the hot Claude Code repos this week aren't about making the agent more powerful. They're about making it accountable. As the agents get more capable, the scarce thing becomes trust, and trust comes from structure.
Repo: github.com/Chachamaru127/claude-code-harness
← Back to all articles
The whole thing is a discipline, not a model. It forces every piece of work through a fixed loop, plan, then work, then review, then release, exposed as five commands: harness-plan, harness-work, harness-review, harness-release, harness-sync. You describe what you want, it drafts a spec.md and a Plans.md with scope, acceptance criteria, and explicit stop conditions, and nothing gets built until you approve that contract. Implementation stays inside the approved boundary. Review runs as a separate step from coding so the thing that wrote the code isn't the thing grading it. Release packages only the verified evidence.
If that sounds bureaucratic, that's the trade. The reason agents drift is that we let them hold the spec, the implementation, and the judgment all in one head. The harness rips those apart and puts a human approval gate between each one. It's the same instinct as code review and CI, just aimed at an agent instead of a junior dev.
It's MIT, it runs on Claude Code 2.1 and up, and it's already deep into the four-point releases, so this isn't a weekend toy. The interesting signal is that the hot Claude Code repos this week aren't about making the agent more powerful. They're about making it accountable. As the agents get more capable, the scarce thing becomes trust, and trust comes from structure.
Repo: github.com/Chachamaru127/claude-code-harness
Comments