Saturday Jun 6
RESEARCHAnthropic

AI Builds Itself, Wants An Off Switch

6JUN
80% STOP

Claude now writes 80% of Anthropic's production code. Fifteen months ago it was nearly zero; today each engineer ships 8x more code. Anthropic published a paper calling for a verifiable global AI pause button.

Claude Code shipped in February 2025. By May 2026, it writes more than 80% of every line merged into Anthropic's own production systems.

Engineers still choose the work, review changes, and decide what merges. But the volume shifted. One engineer used Claude to ship 800+ fixes and cut an API error rate by 1,000x. Claude's success rate on the hardest open-ended internal tasks hit 76% in May - a 50-point gain in six months.

Anthropic's Institute paper doesn't claim recursive self-improvement is here. It maps the path to it - and calls for a global pause mechanism before that line is crossed. The lab making the case for a pause button is the same lab demonstrating the loop in production.

full brief & sources

⚡ Why this matters

  • AI is now training AI at scale inside the world's leading safety lab. The feedback loop is live, not theoretical.
  • The bottleneck shifted from writing code to reviewing it. Human judgment is still required, but time pressure is intensifying fast.
  • Anthropic calling for a global pause button while demonstrating the loop in production is a rare moment of institutional self-awareness. Read it.

🔍 What happened

  • 80%+ of code merged into Anthropic's production systems in May 2026 was authored by Claude.
  • Claude Code launched Feb 2025. Code authorship share went from low single digits to 80%+ in ~15 months.
  • Engineers at Anthropic now ship 8x more code per quarter than in 2024.
  • Claude's success rate on the hardest open-ended internal engineering tasks: 76% in May 2026 (was ~26% in September 2025).
  • Anthropic Institute paper maps the path to full recursive self-improvement and calls for a 'verifiable global pause mechanism.'
  • A parallel project: 9 Claude agents ran an AI safety research task end-to-end, recovering 97% of performance on a benchmark over 800 compute-hours.

💬 Smart takes

  • Anthropic engineer (quoted in paper): 'What I'm offering is seeing the bigger picture beyond the immediate task.' - The human role in review, not generation.
  • CrowdStrike's Elia Zaitsev (via Glasswing report): 'What once took months now happens in minutes.' - This applies to adversarial code generation too.
  • Skeptic - WinBuzzer / Markus Kasanmascheff: The 80% authorship number and 76% success rate show Claude's contribution inside Anthropic's workflow. They don't show that generated changes are automatically safe, maintainable, or ready to merge.

🧭 Where this goes

  1. LikelyThe 80% figure crosses 90% for Anthropic by Q1 2027 as Claude Code matures.
  2. LikelyEnterprise teams face audit-trail and review-gate requirements before AI-authored code can reach production at scale.
  3. PossibleAnthropic's call for a global pause mechanism gets adopted as a formal policy proposal by a major government body within 12 months.
  4. PossibleA competitor publishes similar code-authorship data - signaling industry-wide acceptance that AI builds AI.
  5. Wild CardThe first major production outage traced to unreviewed AI-authored code triggers regulatory action on AI coding tools.

🥄 The Spoon Take

Anthropic built a safety lab to prevent AI from running away from humans. Now Claude writes 80% of Anthropic's code. The loop is live. They know it. That's what the pause-button paper is. It's not a theoretical concern - it's a CYA memo written in real time while the loop runs in production.

🤔 Pushback

80% code authorship at one lab doesn't generalize - Anthropic's workflow is unusually Claude-optimized, and most enterprise teams are nowhere close to this automation density.

·Anthropic Institute - Recursive self-improvement paper·WinBuzzer - Claude writes 80% of code·The Decoder - Claude 90% code, pause button