Google Releases Modern Web Guidance 1.0 For AI Coding Agents
Core Features and Governance
The 1.0 release introduces an audit function for existing code. The system analyzes HTML, CSS, and JavaScript files. It also checks package.json to identify safe update opportunities. This audit relies on static analysis of limited code patterns. The guidance covers hundreds of verified AI knowledge gaps. Each guide is reviewed and approved by a domain expert. Accessibility-focused guides undergo specialized review by assistive technology users. A governance document was uploaded to the project's GitHub repository on November 9. It establishes six council seats. Two belong to Chrome, one each to Microsoft Edge and Mozilla, and two to independent web experts. One independent seat remains vacant. Apple holds no seat on this council.
Breaking news:
Performance Metrics and Evaluation
Google published an evaluation panel comparing AI coding agents from Anthropic, OpenAI, and Google. These agents perform identical daily tasks with and without the guidance. A real browser validates the DOM, styles, and source code of each finished application. Guided runs passed 88% of automated checks. Unguided runs passed only 56%. These results come from overnight runs completed on November 8. They cover 199 to 215 tasks per agent. The panel offers limited accessibility checks, such as keyboard support and form labels. Task instructions do not explicitly mention the guidance. Agents must decide when a guide applies. Recent suites show guided runs were slower and consumed more tokens.
Accessibility and User Feedback
Accessibility consultant Adrian Roselli evaluated the preview version in May. He tested Google's Antigravity AI coding tool with instructions for an animated accordion. The output did not meet WCAG standards. The animation also failed in Firefox. He noted no guarantee that agents will strictly follow the guidelines. These observations refer to the preview, not v1.0. Japanese content platform note inc. tested the new guidelines using two coding agents. They completed six UI tasks based on 180 test runs on cloned production code. Scores for both agents improved after installing the skill, even when not named in instructions. Agent One rose from 60.8% to 76.5%, then to 94.0%. Agent Two recorded 75.8%, 86.2%, and 90.0% under the same conditions.
Real-World Implementation
MetaMask implements the guidance as part of an automated review process for pull requests. This applies to its open-source browser extension. By October 21, the bot identified 102 issues across 57 pull requests. Engineers fully resolved 43 issues and partially resolved three. Another 16 were rejected or false positives. The remaining 20 stayed unresolved, with unknown final outcomes.
Why This Step Matters
Modern Web Guidance offers a zero-cost method for development teams. It encourages AI agents to use native browser capabilities. This applies to legacy codebases as well. Most published results rely on Google's internal testing. They also include client tests highlighted in the official publication.
More stories: