Claude Fable is relentlessly proactive
Points and comments are a snapshot, not live.
Claude Fable autonomously created custom tools and modified application code to debug a CSS scrollbar bug, costing $12 in tokens.
Simon Willison documented Claude Fable 5 debugging a horizontal scrollbar in Datasette Agent's modal dialog. Given only a screenshot and instruction to check dependencies, Fable proactively opened browsers, created test HTML files, built a custom Python CORS server to capture JSON measurements, injected JavaScript into application templates to trigger the bug, and navigated Safari's DOM to extract data. It used PyObjC to query macOS window information and screencapture to grab targeted screenshots. Fable eventually hit a guardrail and downgraded to Opus, which completed the fix: a single-line CSS change (overflow-x: hidden). The session consumed $12.11 in tokens. Willison notes this demonstrates both the capability and security risk of coding agents operating outside sandboxes, as this proactivity could enable data exfiltration under malicious prompts.
What commenters are saying
Commenters focused on token efficiency rather than the security angle. The dominant response noted the disproportionate cost for a trivial fix: the linked diff shows one line changed, a task a frontend developer would solve in seconds using browser dev tools. Several pointed out that Fable's elaborate reverse-engineering process was unnecessary since finding overflow-x: hidden is straightforward. A secondary discussion emerged on AI productivity claims versus actual shipping velocity; skeptics argued that despite 10x productivity claims, major output hasn't materialized. One commenter defended AI as a tool requiring skilled direction, not a replacement for thinking.
Willison replied that he pays his own subscription costs ($100-200/month) and declines free tiers to avoid conflicts, though he retains GitHub's free tier for open source maintainers.