· 3 min read
Agents Write Yesterday's Web
Coding agents default to the web of old tutorials. Browser makers now hand them current guidance while they work, and every platform owner should do the same.
Narrated by an AI clone of my voice.
Researched and drafted with Claude; views are my own.
A coding agent writes the web it was trained on, and that web is mostly old tutorials: dark mode hung on a class or attribute that JavaScript sets, a modal built from divs, a tooltip positioned by a script. Browsers now ship native answers to most of these. The agent has usually seen the modern answer too; it’s just outvoted by ten years of tutorials.
That’s not a quality problem in one project. It’s a distribution problem, and it needs a distribution fix. Teams that write code with agents now ship yesterday’s patterns at scale, and the review that might catch them is increasingly done by another agent with the same habits. Newer models don’t solve it, because volume beats recency, and neither does web search: an agent searches when it’s unsure, and an old pattern doesn’t feel unsure.
Docs have a new reader
Chrome previewed Modern Web Guidance at Google I/O in May, and on 10 November it reached version 1.0, with Firefox and Edge joining Chrome to make it a cross-browser effort. It’s a set of skills a coding agent loads while it works. Each covers a common task, such as a dialog, a form, a tooltip or dark mode, with the modern approach, its pitfalls and its fallbacks, tied to Baseline so the agent knows what is safe for the browsers you support.
Much of the knowledge is on MDN and web.dev already. What’s new is that it’s curated and tested for agents, and delivered to them. Documentation used to wait for a developer to go and read it; now the people who build browsers put their expertise into the agent’s context at the moment the code is written. By Chrome’s own checks, agents equipped with it improved their adherence to modern best practices by 37 percentage points on average in early results, with the caveat that results vary by project, model and prompt. One outside team, note inc., reports pass rates rising from 61% to 94% on its production codebase.
The same move as MCP
This is the argument I’ve been making about airlines, pointed at the web itself. An agent that has to guess its way through a website screenshots it and fumbles the date picker; one that is given tools through MCP or WebMCP books the flight. Tools and knowledge are different things, but the move is the same: the owner hands the agent what it would otherwise have to guess.
That suggests a rule for anyone who owns a platform, an API or a design system: the most frequent reader of your docs is becoming an agent. Ship them in a form an agent loads, not only as a page a person might visit. Concretely, that’s a skill or an MCP server an agent can search and pull from, plus an eval that runs the same task with and without it, the way Chrome tests each guide. For an airline, it’s an MCP server describing what the airline can do. For browsers, it’s a skill describing what the web can do now.
Tried on this site
I ran it on this site this week. The theme toggle, written by a coding agent during the August rebuild, had four quiet bugs: it pinned your choice forever, had no dark mode without JavaScript, never declared both schemes before the stylesheet loaded, and ignored other tabs. One scan flagged the pattern, and the guide behind it showed all four. The fix was written by an agent too, this time with the guide loaded.
Nothing dramatic: these are the bugs that ship for years. An agent will look for them now, as long as the people who own the platform tell it what to look for.