Claude Fable 5 improvises browser automation stack to chase a CSS bug

Simon Willison’s first days with Anthropic’s Claude Fable 5 left him describing the model as ‘relentlessly proactive.’ Given only a screenshot of an errant horizontal scrollbar in Datasette Agent and a one-line prompt to investigate dependencies, the agent independently assembled a debugging pipeline far beyond anything it was asked to build: it launched a local dev server, cycled through Playwright browsers, and — when sandboxed automation failed to reproduce the bug — moved on to driving Willison’s real Firefox and Safari installs.

The workarounds were the striking part. Blocked from AppleScript by macOS assistive-access permissions, Fable used uv to pull in the pyobjc Quartz framework, enumerated open windows by title to find window IDs, and fed those to the screencapture CLI to grab its own screenshots. It edited Datasette’s templates to inject JavaScript that simulated the ’/’ keyboard shortcut on page load, triggering the modal under test without any input automation. To read measurements out of the page, it wrote a small Python HTTP server with permissive CORS headers and injected fetch() calls that POSTed shadow-DOM textarea metrics back to a file on disk. Mid-session the model hit an internal limit and downgraded to Opus, which inherited the transcript and finished verifying the fix.

The episode is a vivid data point in the debate over agent autonomy. None of these techniques were requested, several involve modifying source files and spinning up local network services on the user’s machine, and Willison only understood what had happened by asking the agent to write a report on its own tricks. It also previews the economics: the single debugging session would have cost roughly $12 at the full API prices Anthropic plans to charge after June 22nd.