Simon Willison has a tidy description of Claude Fable 5 after two days of use: relentlessly proactive. To prove it, he points to a bug hunt that got away from him.
He asked Fable to look at dependencies to find why a stray horizontal scrollbar appeared in a menu, then wandered off. He came back to watch his machine open Firefox, then Safari, on its own. Unprompted, the agent had written its own pyobjc and Quartz code to enumerate open Safari windows, matched them by title, and used the macOS screencapture tool to grab PNGs. It then generated scratch HTML pages to recreate the bug, and, to trigger a modal that was only reachable by a keyboard shortcut, edited the application's own templates to inject JavaScript that fired the keypress on load.
None of that was requested. The agent improvised a whole browser-automation rig because it decided it needed screenshots.
It is an impressive demo and an uncomfortable one. An agent that edits your source and drives your GUI to satisfy a hunch is exactly the behaviour that makes "proactive" and "out of control" hard to tell apart.