Google Ships Modern Web Guidance 1.0 to Teach AI Coding Agents the Modern Web

Google has released Modern Web Guidance 1.0, a free skill set that teaches AI coding agents how to build for the modern web, and its early evals show guided agents passing 88% of automated checks against 56% for unguided agents. Announced on 10 November 2026 by Philip Walton, a senior staff developer relations engineer at Google, on the Chrome for Developers blog, the toolkit embeds current web-platform expertise into coding agents so they write modern code instead of falling back on outdated JavaScript-heavy patterns. Modern Web Guidance is a skill set — a curated library of guidance that an AI coding agent can consult — not a new AI model or a new agent of its own.

Why AI coding agents need web guidance

AI coding agents such as Claude Code, Cursor, Copilot, Codex and Antigravity have become the way many developers write code, but the models behind them were trained on years of web history, including patterns the platform has long since outgrown. Left to their own defaults, an agent can reach for a JavaScript library or a hand-rolled workaround when the browser now offers a native feature that does the same job faster and with less code. Modern Web Guidance is Google's answer to that gap: it gives the agent a current, curated source of web-platform knowledge, including browser-compatibility data from Baseline, so its suggestions favour native browser features over unnecessary JavaScript and dependencies. Baseline records when a feature is supported across the major browsers, so an agent consulting it can tell whether a feature is safe to use today.

What is in Modern Web Guidance 1.0

The headline figure is the size of the library: Chrome describes Modern Web Guidance as containing more than 200 guides, each covering a modern web feature or a real-world use case. There is a second, separate count to be careful with. The project repository's README counts 250 modern web features and 203 real-world use cases, broken down as 140 CSS and layout items, 74 JavaScript and API items, and 36 HTML and DOM items. Those are two different ways of counting the same body of work, and they should not be merged into a claim of "250 guides". The guides themselves are the teaching material: short, practical instructions an agent reads before it writes or edits code, covering how a feature works, when to use it, and what to avoid.

The toolkit is free and open: software under Apache 2.0, guide content under CC-BY 4.0, at github.com/GoogleChrome/modern-web-guidance. Install with npx modern-web-guidance@latest install, which runs an interactive wizard, or through agent plugin marketplaces. Preview users from Google I/O in May 2026 can upgrade with npx modern-web-guidance@latest update.

The new modernisation workflow

The biggest addition in version 1.0 is a modernisation workflow, effectively an audit for existing code. It statically scans a project's HTML, CSS, JavaScript and package.json files, looking for safe upgrade opportunities — places where an older pattern can be replaced with a modern native feature without changing what the site does. The agent then walks the developer through the changes one at a time, rather than rewriting the codebase in a single pass. Two limits matter. First, the audit covers only a limited set of patterns in version 1.0, roughly 20% of the guide catalogue, so it will not find every possible upgrade. Second, there is no automatic fix mode: nothing is changed without the developer's involvement, which keeps the workflow safe for production code but means it is a guided review rather than a one-click modernisation.

A web developer's desk with code on a large monitor and a browser window beside it.
A web developer's desk with code on a large monitor and a browser window beside it.


Pairing with Chrome DevTools for agents

Google also recommends pairing Modern Web Guidance with Chrome DevTools for agents, which lets an AI agent run and test the code it writes inside a real browser. Guidance tells the agent what to write; DevTools lets it check whether the result actually works. Together they close a loop that text-only agents cannot: write modern code, run it in Chrome, see the result, and fix what breaks.

Who governs it — and why that matters

Modern Web Guidance is deliberately cross-browser, and its governance reflects that. Chrome holds two seats, Microsoft Edge holds one, Mozilla holds one, and two independent web-platform experts hold seats — one of which was still open when the governance document was uploaded on 9 November 2026. Apple holds no seat. The project is stewarded together with MDN, the web documentation project developers have trusted for years. That structure matters because a guide that only reflected Chrome's view of the web would be of limited use to developers shipping to every browser. Cross-browser governance is Google's argument that the guidance is a shared web-platform resource rather than a Chrome marketing exercise, and the involvement of Edge and Mozilla gives that claim real weight.

What the evals actually show

Google publishes a public nightly eval dashboard that runs the same tasks through agents from Anthropic, OpenAI and Google, with and without Modern Web Guidance, so anyone can see the difference the guidance makes. The headline numbers come from nightly runs up to 8 November 2026: guided agents passed 88% of automated checks, against 56% for unguided agents. Those runs used the preview-era guidance, so they should be read as evidence from the preview period, not as version 1.0 results — an important distinction, since the 1.0 library and workflow are new. Chrome also cites two further vendor-reported figures from its own guide: pass rates on 19 browser checks rose from 61% to 94% across 180 runs, and adherence to the guidance improved by an average of 37 percentage points. Both are Google's own figures and should be treated as vendor-reported, not independent benchmarks, but the direction is consistent: agents with the guidance write code that matches modern web practice far more often than agents without it.

Close-up of hands typing HTML and CSS code on a laptop keyboard.
Close-up of hands typing HTML and CSS code on a laptop keyboard.


How it fits the agent-coding landscape

Modern Web Guidance arrives at a moment when AI coding agents are becoming the default way software is written, and the quality of their output depends heavily on what they are taught. Anthropic's Claude Duet and OpenAI's OpenAI Decisions API show how quickly agent capabilities are expanding, and Google's own Gemini edit files feature brings agent coding to everyday users. Modern Web Guidance is a different kind of contribution: instead of a smarter agent, it is a smarter curriculum. The same agent, given better teaching material, writes better web code. That framing also explains why Google made it free and open rather than tying it to a single product — the value grows as more agents, from more vendors, adopt it, and the cross-browser governance is designed to make that adoption easy to trust.

What developers should do now

For a developer already using an AI coding agent, the practical step is simple: run the install command and let the wizard set it up, then try the modernisation workflow on a small project before pointing it at production code. Expect the audit to find only a fraction of possible upgrades in this first version, and expect to review each suggested change rather than accepting them blindly. The broader question is whether other parts of the software stack get the same treatment: if a curated, cross-vendor skill set can lift an agent's web code from 56% to 88% on automated checks, the same approach could apply to accessibility, security, or performance.

A smartphone and a laptop side by side showing the same responsive website.
A smartphone and a laptop side by side showing the same responsive website.


Key Takeaways

  • Modern Web Guidance 1.0, released 10 November 2026, is a free skill set that teaches AI coding agents modern web features — not a new model or agent.
  • Install with npx modern-web-guidance@latest install; the project is Apache 2.0 software with CC-BY 4.0 guide content, stewarded with MDN.
  • Version 1.0 adds a modernisation audit of HTML, CSS, JS and package.json, covering about 20% of the guide catalogue with no automatic fix mode.
  • Preview-era nightly evals to 8 November show guided agents passing 88% of automated checks versus 56% unguided — not version 1.0 results.
  • Governance is cross-browser: Chrome (2 seats), Edge (1), Mozilla (1) and two independent experts; Apple holds no seat.
  • Chrome's headline count is 200+ guides; the repo's separate count of 250 features and 203 use cases should not be conflated with it.

Conclusion

Modern Web Guidance 1.0 is a modest tool with an ambitious idea behind it: that AI coding agents should be taught the current web, not left to guess from old training data. The 88% versus 56% eval gap from the preview period suggests the teaching works, and the open licence, cross-browser governance and public eval dashboard give developers good reason to try it. Run the install command, test the modernisation workflow on a small project, and judge the results in your own browser before rolling it out to a whole codebase.

Sources

Comments

Post a Comment