Claude Code AFK Auto-Continue: Anatomy of a Misfeature

On Canada Day 2026, Anthropic silently shipped Claude Code v2.1.198 with an 'efficiency bypass' that gives agents a 60-second window after asking for human input. If you miss the window, Claude Code continues on its own with 'best judgment'. No changelog entry mentioned this new default.

What Happened

• Version 2.1.198 introduced the behavior. User Olaf Alders reported the session log: ● Claude asked: ... ● No response after 60s — continued without an answer ● The user stepped away. I'll proceed with best judgment.
• Issue #73125 was filed by Aleksey Nogin on 2026-07-02 02:54 UTC.
• A commenter discovered the escape hatch: CLAUDE_AFK_TIMEOUT_MS. This env var was not documented in any release notes.
• Version 2.1.199 shipped 24 changelog entries while the issue was open — still no mention of the feature.
• Version 2.1.200 on 2026-07-03 reversed the behavior. Again, the only public trace was a changelog commit with no code diff visible.

Key Concerns Raised

• Users running multiple agents cannot watch all input windows simultaneously.
• If two agents ask for input in the same 60-second window, both may auto-continue with wrong decisions.
• Token waste from an agent continuing on a wrong path before you can intervene.
• Deployment safety: an auto-continue could push a bad deployment.
• Trust: a silently shipped default with no off switch documentation erodes confidence.

Questions That Remain Unanswered

The author asks whether a human actually reviewed, signed off, and merged this feature — or if an AI agent shipped and documented it without human gating. The lack of changelog entry and the speed of the silent revert suggest process failures in Anthropic's release pipeline.