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.
📖 Read the full source: HN LLM Tools
👀 See Also
AI-generated code volume is overwhelming senior engineers, study shows
AI users merge 98% more pull requests with AI assistance, but senior engineers report increased cognitive load and burnout. Research shows defect detection drops from 87% for PRs under 100 lines to 28% for PRs over 1,000 lines.
PwC 2026 CEO Survey: 56% Report Zero Financial Return from AI, Only 12% Succeed
PwC surveyed 4,454 CEOs across 95 countries and found 56% report zero financial impact from AI, while only 12% have successfully used AI to both cut costs and grow revenue. The successful 'Vanguard' companies are 3x more likely to apply AI directly to products and services.
Reddit Discussion Critiques Reactive AI Assistants, Calls for True Proactivity
A Reddit post argues that current AI assistants are reactive by design, waiting for human prompts rather than proactively identifying issues. The author distinguishes between scheduled checks and true contextual awareness, noting that real proactivity requires persistent memory, event-driven triggers, and cross-time reasoning.
RTX 5000 PRO 48GB Delivers 4400 tok/s Precision Caching for Qwen3.6-27B
A first-time PC builder reports 4400 tok/s prompt processing and 80 tok/s generation with Qwen3.6-27B-FP8 full-precision KV cache on a single RTX 5000 Pro 48GB, using vLLM and Claude Code.