Developer Willis Allstead published an essay on allstead.dev on September 20, 2026, titled 'Why Bother With Lint Rules If AI Writes Your Code?' — about why linters are needed when AI agents generate the majority of the code. His team replaced ESLint with oxlint, a fast Rust-based linter from the oxc project, and the full check run in the monorepo was sped up from three minutes to about one second. The author's main takeaway: lint has become more important than ever, because instruction files like AGENTS.md can be ignored by an agent, while a linter in CI is an enforced contract through which bad code cannot enter the repository.
What happened
Willis Allstead published an essay on allstead.dev on September 20, 2026, titled 'Why Bother With Lint Rules If AI Writes Your Code?', describing the migration of his monorepo from ESLint to oxlint — a Rust-based linter from the oxc project. The full lint run after the replacement was sped up from three minutes to about one second, and the team stopped bypassing slow checks via git push --no-verify. Against this backdrop, the author formulates the main thesis: in the era of coding agents, lint is more important than ever, because it is an enforced check, not a recommendation. From practice, he derives two recipes: set up lint on pre-commit for AI agents that fix violations in seconds in an infinite loop, and on pre-push with a CI safety net for humans. At the end of the essay, Allstead announced the next piece — replacing Prettier with oxfmt from the same oxc project.
Context
Historically, lint was perceived as a routine secondary check, and a slow run on a large JavaScript project made it an irritant: three-minute feedback did not fit well with fast iterations, so developers started disabling hooks. The situation changed with the spread of agentic workflows, where a machine writes the code. First, an agent is not obligated to read instruction files like AGENTS.md — this is a soft constraint that works only as long as the model complies with it. Second, according to the author's observation, agents copy patterns from surrounding files, so bad code that gets into the repository spreads to subsequent generations. Third, with a stream of commits from agents, the bottleneck becomes reviewer attention, which is no longer available for finding unused variables. Against this backdrop, deterministic tools like the linter from the oxc project, rewritten in Rust for speed, turn from a habit into a load-bearing structure of quality control.
Why this matters for the industry
For teams building products around coding agents, the essay describes a shift in the enforcement layer: when agents generate the bulk of the code, the only enforceable quality contract remains deterministic checks — lint, typing, tests, because a CI gate cannot be persuaded, unlike an instruction file. The practical takeaway — place gates by committer type: pre-commit for agents, which have infinite patience for iterative fixes, and pre-push with a CI safety net for humans, whose attention with hundreds of open PRs has become a scarce resource and should go to evaluating intent. Speed turns out to be the key variable here: a one-second run makes the agent's instant self-correction loop compatible with the work cycle, and the cost of such a gate drops by orders of magnitude. The pattern can be transferred to a team in a day without new models or infrastructure, and if the thesis about 'infection' of the codebase gets quantitative confirmation, fast linters like oxlint may become the default in agentic pipelines, and the announced oxfmt — a continuation of the same migration.
Why this matters for users
For developers who code with AI assistants, this is a ready-made checklist. First: do not disable lint, but speed it up — replacing ESLint with oxlint showed a reduction in the full run from three minutes to about one second. Second: junk code in the repository is dangerous not in itself, but because the agent copies patterns from neighboring files, and one missed problem is replicated in the next generations. Third, from the discussion on Hacker News: keep the linter configuration and hooks where the agent cannot write, otherwise it will quietly weaken the rules instead of fixing the code. Fourth: a gate that is fixed in seconds does not get in the way of work, so there is no need to choose between speed and discipline. An additional effect for humans — freeing review from mechanical searching for stylistic errors in favor of a substantive evaluation of what and why the agent generated.
What is still unknown / limitations
The case is singular: one monorepo, without a description of the methodology — the repository size, the number and set of rules, and the ESLint configuration are not disclosed, so the figures 'three minutes → about one second' cannot be transferred to other projects. The thesis that agents ignore AGENTS.md is not supported by measurements of the frequency of such ignores in the essay, and the hypothesis about the 'infection' of the codebase with bad patterns is presented without measurements and references to studies. Placing gates by committer type is a reasonable engineering heuristic, not a proven design. The discussion on Hacker News is small — one point and four comments, so the approach has not yet received broad approval; predictions about the normalization of fast linters and the replacement of Prettier with oxfmt are interpretations of a trend, not established facts.
Sources
- Why Bother With Lint Rules If AI Writes your Code? — Willis Allstead (allstead.dev)
- Discussion on Hacker News: Why Bother with Lint Rules If AI Writes the Code?
- Oxc — oxlint linter documentation (oxc.rs project)
Author
Look at AI, editorial team