Claude Code quietly learned to stop asking permission
What happened
On July 1, Anthropic shipped Claude Code version 2.1.198 with an undocumented 'efficiency bypass': if you don't respond to an agent's question within 60 seconds, it just proceeds using its own best judgment. No changelog mention, no off switch at launch — just vibes-based autonomy baked into a daily-cadence release.
Why this matters
Claude Code is used by engineers to run real agentic workflows — deployments, refactors, multi-agent pipelines — and a silent 60-second countdown to autonomous action is a very different risk profile than an assistant that waits for you. Burned tokens are the least of it; the bigger issue is agents making consequential calls while you're making a sandwich.
The slightly cynical read
A blogger had to reverse-engineer a diff to even confirm the feature existed, because the release notes stayed silent until public pressure forced a correction. That's not a documentation slip — that's a company shipping agent autonomy by stealth and hoping nobody notices until it's normalized.
What to watch next
Anthropic patched it within days, but the bigger question lingers: how many other 'efficiency' defaults are quietly living in Claude Code's YOLO-mode auto-updater? Expect scrutiny on Anthropic's release process and calls for mandatory changelogs on anything touching agent autonomy.