AI agents write code fast. Then the PR sits in review for days.
Sound familiar?
- It works on your machine. Review finds what the agent missed.
- The PR is 40 files. Your reviewer skims and hopes.
- The reviewer asks why. You can’t explain code your agent wrote.
It’s not the agent’s fault. It just needs a workflow.
So I built harnex. It runs in Claude Code, Codex and OpenCode.
How it works
- You read a plan in 30 seconds and approve it.
- The agent writes a failing test first, then the smallest change that works.
- A validator and a fresh reviewer check the work, up to 3 rounds.
- Your reviewer’s checklist runs before they see the PR.
More risk, more checks: Quick, Standard or Full, picked from the change. Any language, any stack.
What does review catch most often in your agent’s PRs?
The project
See the harnex project page or the code on GitHub.