Let's be precise about the problem, because "agent code is bad" isn't it. Agent code is mostly fine, which is why the slop gets through.
Slop, as I'm defining it: code that is correct, passes review individually, and makes the codebase worse collectively. The specific strains I keep merging and regretting:
// Iterate over the users above a for loop over users. Individually harmless. Two hundred of them and the codebase reads like it's explaining itself to a jury.The insidious part: none of these fail review. Each one survives "is this line wrong?" It's the aggregate that's wrong, and aggregate review isn't a thing.
What has actually moved the needle for people — instruction rules, lint enforcement, review checklists, something else? Bonus points for the style-archaeology problem specifically, because I think it's the one that compounds.
Strong evidence gets promoted into the record above.
On style archaeology: we stopped it with a canon file. docs/canon/ holds one exemplar file per pattern — THE way we do a service, THE way we do error handling — and AGENTS.md says 'when in doubt, imitate canon/, never neighboring files.' The insight is that agents were never choosing our old style; they were choosing the majority style. Canon breaks the tyranny of the corpus. Deprecated patterns stopped spreading within a month, measured by grep.
Adopted this after you mentioned it in another thread. Confirming it works, with one amendment: the canon files must be real files that compile and ship, not documentation samples. Agents detect and discount toy code — imitation quality jumped when we pointed at production files instead of curated snippets.
Defensive ceremony is worse in C++ than anything described here — agents wrap RAII types in null checks and add try/catch around noexcept calls, which is not caution, it's superstition. My fix is blunt: the AGENTS.md line 'every defensive check must name the failure it defends against in a comment, or be deleted.' Forcing the agent to articulate the threat kills 90% of the ceremony, because there usually isn't one.
Half-disagreement worth logging: some 'slop' is real, but I've watched teams use the label to reject unfamiliar-but-fine code, and agent PRs get slop-flagged at a rate human PRs with identical patterns don't. Before adding ceremony, define it: our rule is a slop objection must cite a written convention, or it's a style preference and doesn't block. Kept our review culture from turning into vibes enforcement against the robot.
Strong agree, with a flip side: accessibility slop is the inverse case — agent markup is uniformly worse than our humans' (div soup, missing labels, decorative alt text), and it needed new checks, not existing conventions. The lesson generalizing both our points: audit where agent output actually differs from your baseline instead of assuming, because it differs in both directions.
Sign in to join the discussion, vote, and verify fixes.
Related records
The only thing that durably worked for me: convert taste into lint rules, ruthlessly, even when the rule feels petty.
eslintno-single-use-abstraction isn't a real rule, but max-lines-per-function plus a custom rule flagging exported symbols with one call site gets 80% of it.The general principle: agents respect machine feedback and negotiate with prose feedback. Anything in AGENTS.md is a suggestion; anything in CI is physics.
'Anything in CI is physics' is the thesis of this whole site, honestly. Do you publish the comment-overlap hook anywhere? That's the strain I have no answer for.
Gist linked from my profile. Fair warning: threshold tuning took longer than writing it — 0.6 overlap flagged docstrings, 0.8 missed half the narration. We run 0.72 and review its rejections weekly like the tiny bureaucracy it is.