3 min read
Coding agents write working web code — just not modern web code
Chrome, Firefox, and Edge have shipped structured guidance to fill the gap in what coding agents know about the modern web platform. Here's why that matters for small teams.
By intermac
Coding agents have made it faster than ever to get something working in a browser. The problem is "working" and "good" are not the same thing. Ask a web specialist to review AI-generated frontend code and they will almost always find accessibility gaps, missed native HTML or CSS features, outdated patterns, and security oversights. The code runs — it just isn't modern.
Chrome, Firefox, and Edge have now shipped a concrete answer to this: Modern Web Guidance 1.0. It's a set of machine-readable guides — covering accessibility, performance, UX patterns, and security — that you install directly into your project so your coding agent can reference them. The guides are maintained by browser engineers, web standards specialists, and accessibility experts, and they're tied to Baseline compatibility data so recommendations adjust to your actual browser support requirements.
The gap is real and specific
The areas where models fall short aren't obscure edge cases. According to the project's evaluations, common failures include keyboard focus management, reduced-motion support, safe HTML sanitization, main-thread task scheduling, and passkey handling. These are the fundamentals that separate a polished product from a brittle one.
What makes the problem worse is that AI models tend to solve problems with JavaScript even when HTML or CSS would do it better, with less code and built-in accessibility. The guidance is explicitly designed to steer agents toward native platform features first.
Why this is worth adding to your workflow
The article makes a point worth taking seriously: once every team is using coding agents, raw output speed stops being an advantage. The differentiation moves to quality — usability, accessibility, performance, security. Tools that raise the floor of agent-generated code matter more as the ceiling of raw speed flattens.
The 1.0 release also adds a modernization workflow that scans your existing HTML, CSS, JavaScript, and package.json for safe upgrade opportunities. That's the more valuable feature for most teams. The majority of web development happens on existing codebases full of polyfills that are no longer needed, workarounds for browser bugs that were fixed years ago, and JavaScript handling tasks that CSS now does natively. Running /modern-web-guidance Modernize the code in this directory gives an agent a guided path through those upgrades rather than leaving it to guess.
Real deployments back up the claims. MetaMask added it as an automated PR review bot — three lines of configuration — and in five weeks it flagged actionable issues in 64 of roughly 480 pull requests, mostly accessibility and HTML semantics problems that linters don't catch. note inc. improved agent pass rates on modern web checks from 61% to 94% across a Next.js production codebase.
The trade-off
Adding any tool to an agentic workflow introduces maintenance surface. The guidance evolves as the platform evolves, so you need to update it. That's the honest cost.
But the alternative — letting agents operate without this context — is not neutral. It means systematic quality gaps that accumulate silently until a user with a screen reader, a slow device, or a low-end network hits them. For small teams especially, those regressions are expensive to find and fix after the fact.
Installing Modern Web Guidance is a one-command decision:
npx modern-web-guidance@latest installIt's also available directly in Claude Code, Cursor, Copilot, Codex, and several other agent environments. For anyone building web products with AI-assisted coding, it belongs in the standard setup.