Dev.to AI 🤖 Ai 👁 0 📖 2 min read

After Months of Use, Our Agent Passes Our Lint Rules

For the first month our coding agent opened pull requests, CI was the embarrassing part. Not logic bugs — lint. Import order, naming style, bare excepts, missing docstrings exactly where they should be. Our repo has ESLi

For the first month our coding agent opened pull requests, CI was the embarrassing part. Not logic bugs — lint. Import order, naming style, bare excepts, missing docstrings exactly where they should be. Our repo has ESLint, Ruff and pre-commit wired up, and human code goes through that gate, so agent code did too. Every red run said the same five things, like reviewing some intern's week one.

Nobody trained it, and nobody changed the config, but somewhere around the third month the style converged.

The interesting part is where the change came from. We run our agents on Octo, an open-source workspace for human-agent collaboration, and the platform has an experience layer: every rejection and every review preference — "we wrap IO errors like this in this repo" — gets distilled into a rule with a scope, a source and a confidence, then injected before the agent picks up its next task in that repository. The lint failures and review comments from month one didn't evaporate. They were the cheapest training data we had, sitting in our own pipeline the whole time.

Two examples. Bare except is caught by the linter, but after getting blocked twice, the agent went beyond fixing the pattern — it started wrapping external calls in the exception types this project already uses. Docstrings on exported functions are a stricter case: that rule exists only in review habit, nowhere in config. Three rejections later the agent wrote them proactively, matching the comment language of the file it was editing. What a linter can check is the floor; the part above the floor was beaten in through reviews.

We track two CI numbers instead of vibes. First-pass rate: the share of PRs where lint runs clean on the very first attempt. It climbed steadily from month one to month three. Repeated-error count: the same rule firing a second time in the same repo. This one drops faster, and honestly it's the more meaningful number — a first-time failure is a model problem, a repeat failure is a memory problem. When repeats stop, rejections are actually being remembered.

Boundaries worth stating. Passing lint means fitting into the repo, not writing good code — reviewers still review logic, and an experience layer will not turn 60 into 90. Convergence has a side effect too: the agent starts writing like the team, including the team's bad habits. If a rule is wrong, it learns the wrongness quickly. So every few weeks we read through the high-frequency experience entries and decide whether to change the rule or correct the agent. With that review loop in place, accumulated experience stays an asset; without it, it becomes debt.

The quiet pleasure of using this kind of platform for months: what grows over time isn't the model, it's how well your own codebase is known. The experience system lives in our open-source repo if you want to see how entries are structured and when they get injected: https://github.com/Mininglamp-AI/Octo — a star helps while you're there.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.