Anthropic warns AI development is becoming increasingly automated

Anthropic is moving more of its AI development workflow to Claude, from coding assistance to autonomous agents that can run code, delegate work and execute research loops. The company frames the shift as early evidence toward recursive self-improvement, where an AI system could eventually design and develop its own successor, while stressing that such an outcome is not inevitable.

Internal data shows Claude now writes a large share of Anthropic’s software. As of May 2026, more than 80% of code merged into Anthropic’s codebase was authored by Claude, compared with low single digits before Claude Code launched in research preview in February 2025. Lines of code merged per engineer per day were flat through 2021-2024, then rose in 2025 and again in 2026; in the second quarter of 2026, the typical engineer merged 8× as much code per day as in 2024.

Claude’s capabilities are also improving on open-ended and research tasks. On the most open-ended coding tasks, its success rate reached 76% in May 2026, up 50 percentage points in six months. In a fixed experiment-optimization test, Claude Opus 4 averaged a ~3x speedup in May 2025, while Claude Mythos Preview reached ~52x by April 2026; a skilled human researcher would need four to eight hours to reach 4x.

The central remaining gap is judgment: choosing goals, deciding which experiments matter and validating results. Anthropic outlines futures ranging from capability gains stalling to AI labs seeing compounding efficiency improvements to full recursive self-improvement, and argues that any slowdown or pause would need credible global coordination and verification across frontier developers.