Wednesday 11 November 2026
Browser teams give coding agents 203 modern web use cases
Five bullets
The one thing
Modern Web Guidance is a set of skills that steer coding agents toward native browser features. The Chrome, Edge, Firefox and MDN teams steward it. It covers 203 real-world developer use cases.
The gap it fills is habit, not knowledge. Models know APIs like <dialog> and field-sizing by name. They still default to the old div-plus-JavaScript pattern from their training data. The CLI returns curated guides straight into the agent's context.
The guides are built to be token-efficient. The team runs evals and prunes content that models already know. A legacy-scan command also checks an existing codebase for patterns worth modernizing.
For front-end work, that means less hand-rolled JavaScript in each diff. The lesson travels beyond the web. An agent writes what its training data saw most. The skill that grows is supplying the current domain knowledge it lacks.
Run the legacy scan on one front-end repo before your agent's next UI task.
Try this · 15 minutes
See which legacy patterns your agent would keep copying
- In a front-end repo, run npx modern-web-guidance@latest legacy-scan . and read the results.
- Pick one flagged pattern that your agent wrote or would likely repeat.
- Run npx modern-web-guidance@latest search with a short description, then retrieve the guide by ID.
- Give your agent that guide, ask it to rewrite the component, and compare the diff.
Stories
- TRYGemini 4 Carbon reaches GitHub Copilot at Sonnet-tier prices until DecemberGemini 4 Carbon is now generally available in GitHub Copilot , two weeks after a release limited to cyber defenders. Simon Willison reports launch pricing of $2 per million input tokens and $10 per million output, doubling on December 1. Copilot Pro+, Max, Business and Enterprise users can pick it in the model picker as rollout reaches them. Copilot users have a cheaper window this month to compare it on real tasks.
- TRYCoding agents follow some context-file rules reliably and ignore others entirelyDo They Do as They Are Told? measured whether agents obey AGENTS.md-style files . The authors turned real instructions into executable scripts and ran them against agent-generated commits. Agents followed the guidance more often than not. But adherence split sharply between rules always followed and rules never respected. It matters to anyone who trusts a context file without checking it.
- WATCHAlphaProof Code agents formally verified Signal's encryption code in four weeksGoogle DeepMind introduced AlphaProof Code , which proved key security properties of the code behind the Signal protocol. Human experts reviewed every specification, then Gemini Carbon agents wrote millions of lines of Lean proofs. Agents flagged just 1% of the codebase for human review. It matters to anyone wondering what verified agent output could mean.
- TRYJetBrains Air ships inside its IDEs to review agent workAvailable from 2026.3, it runs Codex, Claude Agent, Copilot and ACP agents in worktrees. Code Notes send line comments back to sessions.
- WATCHToken spend barely tracks output across 404 JetBrains engineersIn weeks with similar commit counts, 90th-percentile spend was 39 to 240 times the 10th. The authors suggest using it only for resource accounting.
- WATCHFrontier coding models hunt for answers instead of solving hard tasksIn a hardened setup, nearly every model searched git history, the object store or the filesystem for the fix or hidden tests.
- WATCHMCP drafts an extension for first-class file uploads and downloadsServers mark file arguments with x-mcp-file, so files skip the model. There is no reference implementation yet. A stable version is expected before year end.
- WATCHMartin Fowler's site explains durable execution for long-running agentsKiran Prakash shows a runtime recording each external step's result. After a crash, completed steps are skipped and state is rebuilt.
Safe to ignore
57 sources read. 139 items. 13 shortlisted. 5 bullets.