Google's Modern Web Guidance 1.0 Wants Your AI Coding Agent to Stop Writing 2019 Code

Nothing here is sponsored — no affiliate links on this page. Every fact below comes from Google's announcement of November 10, 2026, as reported by Search Engine Journal. We have not tested Modern Web Guidance hands-on — this is launch-day researched analysis, not hands-on testing.

If you have ever watched an AI coding agent confidently write code that would have been fine five years ago — a <div> button instead of a real <button>, a hand-rolled modal instead of the <dialog> element, a polyfill for something every browser has shipped for years — you know the problem Google is trying to solve. The model is not stupid. It was trained on the whole history of the web, and the whole history of the web is mostly old code.

What it is

Modern Web Guidance is a set of guides — a skill — that you install into your coding agent. It covers what Google calls "hundreds of verified AI knowledge gaps": the places where models reliably reach for an outdated pattern because the outdated pattern dominates their training data. Every guide is reviewed and approved by a domain tech lead at Google, and guides for user-facing features get a dedicated accessibility review by experts who use assistive technologies every day. That last detail matters more than it sounds: an agent that picks the modern element usually also picks the accessible one, and Google is checking that claim with people who actually depend on it.

Installing it is one command in a project directory:

npx modern-web-guidance@latest install

It is also available through the plugin marketplaces of most popular coding agents. If you tried the preview, update with npx modern-web-guidance@latest update.

The new thing in 1.0: an audit for code you already have

The headline feature of 1.0 is not for new code. It is an audit for existing code. Point it at a project and it walks your HTML, CSS, JavaScript and package.json looking for safe modernisations — places where an older pattern can be replaced with a current browser feature without changing behaviour — and then walks the agent through making the changes.

The honest caveat, straight from Google: this first iteration uses static analysis over a limited set of patterns. It will not find everything, and it will not reason deeply about your code yet. That second part is explicitly on the roadmap — Google plans to expand auditing to cases where the agent reasons about the code it is auditing. For now, treat the audit as a solid first pass over the obvious wins, not a full modernisation review.

Firefox and Edge are joining the project, which is worth noting: this is not a Chrome-only play. The guidance is being uploaded to GitHub — it went up on November 9 — under a governance model with six owner seats: two from Chrome, one from Microsoft Edge, one from Mozilla, and two independent web platform experts. One independent seat is still open. And, in a detail that tells you something about the politics of the web platform, no seat belongs to Apple.

The numbers: 88% vs 56% — with a real cost attached

Google runs a public eval dashboard where coding agents from Anthropic, OpenAI and Google run the same tasks every night, with and without the guidance. A real browser checks the DOM, the styles and the source code. In the nightly runs completed November 8 — 199 to 215 tasks per agent — guided runs passed 88% of automated checks, against 56% unguided.

That is a big gap, and it is the number Google will lead with. But Google also reports the cost, and it deserves credit for doing so: guided runs were slower and burned more tokens. That is the trade, stated plainly. You are buying correctness with time and tokens. For a UI that ships to users, that trade is usually worth it. For a throwaway prototype, it probably is not. The guidance is not free in the sense that matters — your token bill — even though the skill itself costs nothing.

The eval is also Google's own, run on tasks Google designed, and the agents tested are the ones that agreed to be tested. Take the 88% as a strong signal, not a settled fact.

What early users actually saw

Two early results are more convincing than the dashboard, because they come from someone else's codebase.

Japan's note inc. tested two agents on 6 UI tasks across 180 test runs. Agent One went from 60.8% passing with no guidance, to 76.5% with the guidance installed, to 94.0% when the guidance was also named in the prompt. Agent Two went from 75.8% to 86.2% to 90.0%. Two things stand out: installing the skill helps, and telling the agent to use it helps more. The second part is a little embarrassing for the state of agent tooling — a skill that sits there unused is just a file — but it is useful to know. If you install this, name it in your prompt.

MetaMask runs the guidance as an automatic review bot on every pull request. In the first five weeks, through October 21, it found 102 issues across 57 PRs. Of those, 43 were fully resolved, 3 partially, and 16 were declined as false positives. That is a real-world hit rate on a real codebase, with real developers deciding what counted. Roughly 40% of what it flagged got fixed outright; roughly 16% was wrong. That ratio — useful, but noisy enough to need a human in the loop — is about what you should expect from any automated reviewer, and MetaMask's numbers say it plainly.

What to try first

• Install it in one real project, not a toy repo: npx modern-web-guidance@latest install. Then name it in your prompt on the next UI task and compare the output against what the agent produced last week.
• Run the audit on your oldest frontend code. That is where the safe modernisations live — the code written before <dialog>, popover, and the rest landed. Review what it flags before accepting anything; the MetaMask numbers say about one in six flags will be wrong.
• Watch the token bill. The guided runs cost more. If your agent bills per token, decide whether the correctness gain is worth it on every task, or only on the UI work that ships.

What not to do is treat this as a reason to stop reviewing agent output. The guidance makes the agent better at picking modern patterns. It does not make the agent right. The 88% figure still leaves 12% failing automated checks, and automated checks are the easy bar.

The bigger picture

The interesting thing about Modern Web Guidance is not the skill itself. It is what it admits: the models we code with are trained on the past, and the past is full of patterns the platform has moved beyond. The fix is not a bigger model. It is a small, maintained, expert-reviewed layer of current knowledge sitting between the model and your codebase — and, notably, one that Firefox and Edge are joining rather than competing with.

Google's next step is to pair the guidance with Chrome DevTools for agents, so the agent can run and test its own code in the browser. That closes the loop the current version leaves open: guidance tells the agent what to write, DevTools would let it check whether it worked. If that lands, the 88% number should climb — and the token bill will climb with it.

For now, one command, one prompt, one old codebase. That is the whole experiment, and it is free to run.

What we don't know yet

• Independent evals: the 88% vs 56% figures are from Google's own public dashboard, on tasks Google designed. No independent reproduction exists yet.
• How deep the audit goes: the first iteration is static analysis over a limited set of patterns; reasoning-based audits are still on the roadmap.
• Token cost in practice: guided runs cost more, but Google has not published a per-task token figure — your bill depends on your agent and your codebase.
• Apple's position: no governance seat belongs to Apple, and it is not clear whether Safari's team will adopt the guidance.

Sources: Chrome for Developers blog (Philip Walton), November 10, 2026; Search Engine Journal, November 10, 2026; GitHub repo (uploaded November 9, 2026).