Three Gaps in AI Coding Agents

▶ Watch (2:47)

AI coding agents default to legacy patterns. A settings menu built by an agent required 172 KB of minified JavaScript. The same prompt with Modern Web Guidance produced the same UI in under 3 KB — no JavaScript at all. Agents also have knowledge cutoffs. Chrome shipped 50 new features in the past year, but models trained before those releases don’t know them. When asked to use the interest invokers API, one agent output the old attribute name interesttarget instead of the shipped interestfor. The third gap: agents make one-size-fits-all browser support decisions, ignoring Baseline data.

Modern Web Guidance: Vetted Skills for the Platform

▶ Watch (6:29)

Chrome’s early preview of Modern Web Guidance is a skills pack installed via npx modern-web-guidance install. It contains two layers: a dozen high-level discipline guides for performance, accessibility, security, and user experience, plus more than 100 low-level use-case guides. Each guide focuses on a task a developer would prompt for — not on the API itself. For example, “smoothly morph the size and position of elements across navigations” maps to the View Transitions API. The guides are concise, token-efficient, and stored locally.

Two-Layer Implementation with Fallbacks

▶ Watch (11:50)

Each guide starts with the ideal implementation using the most modern solution, ignoring browser support. Then it documents a fallback strategy for features not yet Baseline Widely Available. The fallback includes detailed browser compat data. When a developer sets a Baseline target in agents.md, the agent can skip fallbacks that are already supported. An Electron app targeting Chrome 144+ can ignore fallbacks for scroll-driven animations (shipped in Chrome 115). This tailors the output to the project’s browser requirements instead of taking a lowest-common-denominator approach.

554 Evaluations and Daily Testing

▶ Watch (14:57)

The Chrome team runs 554 distinct eval assertions daily against Gemini 3.1, Claude Opus 4.7, and GPT 5.5. In aggregate, agents with Modern Web Guidance pass measurably more assertions than unguided agents. One eval for a scroll progress indicator checked that the implementation used scroll-driven animations instead of scroll event listeners. Guided agents passed all assertions; unguided agents failed all of them. The team plans to remove guides once unguided performance catches up, but expects new features to keep the pack necessary.

Building a Mobile Email App with Native Feel

▶ Watch (21:43)

Walton live-coded a mobile email app using two branches of the same codebase. The unguided agent added a swipe-to-close gesture by detecting the swipe and calling a close function — the drawer didn’t track the user’s finger. The guided version used scroll-driven animations and scroll snap points. The drawer pinned to the finger, the backdrop faded proportionally, and swipe-to-close worked with native scroll physics on both a Pixel and an iPhone. Guided also added a stack-drill-down transition with parallax and swipe-to-archive on inbox rows.

Notable Quotes

“Why are AI coding agents recommending a solution that requires almost 172 KB of JavaScript when the same exact solution can be built in a way that requires no JavaScript at all and still works across browsers?” Philip Walton · ▶ Watch (1:56)

“If you ask a coding agent to use a feature that it’s never heard of, it’ll usually do a web search. But if you ask it to do something that it thinks it already knows how to do, then it’ll often just reply confidently right away without checking for more up-to-date information.” Philip Walton · ▶ Watch (3:57)

“Guided pass rate the guided agent performs meaningfully and noticeably better than the unguided agent.” Philip Walton · ▶ Watch (15:58)

“I can swipe it to close and you know, I can swipe it like that and that looks great.” Philip Walton · ▶ Watch (26:45)

Key Takeaways

  • Modern Web Guidance cuts agent-generated JavaScript from 172 KB to under 3 KB for typical UI.
  • More than 100 use-case guides are locally stored and semantically matched to developer prompts.
  • Two-layer fallback logic adapts to each project’s Baseline target, avoiding lowest-common-denominator code.
  • A 554-assertion eval suite runs daily across three major models to measure guided vs unguided pass rates.
  • Chrome commits to adding guides for every new web feature before it lands in stable.