Claude Fable will build a browser-automation rig to fix your CSS

Simon Willison walked Claude Fable 5 through a small but stubborn bug in his Datasette Agent project: an unwanted horizontal scrollbar in a modal dialog. What he reports is less about the bug than how the model went about finding it. Without being asked, Fable wrote test HTML files, injected JavaScript into application templates so the modal would open on page load, scripted Safari window selection in PyObjC to filter for windows whose titles contained "textarea", drove screencapture from the command line, and stood up a small Python http.server to receive JSON measurements from the page over CORS. It also went through shadow DOM via Web Components to read textarea properties at runtime.

The number that matters is at the end. Fable used about $12.11 in tokens for the session. The final fix was two lines of CSS.

Willison's reading is that this is what "agentic" looks like in practice now: the model treats the boundary between "tell me what's wrong" and "go set up a measurement rig" as not the user's problem. It is also expensive. Twelve dollars to find a two-line change is fine for a real production bug, less fine as a default loop you leave running on every minor ticket.

Why it matters

If you hand a top-tier model a problem and shell access, it will do things that used to take a senior engineer an afternoon to plan, and it will burn tokens to do them. The practical decision now is setting a budget ceiling for agentic debugging, and deciding which kinds of bugs justify letting the model run unsupervised long enough to build its own browser-automation harness.