When AI Builds Itself: Our progress toward recursive self-improvement

Anthropic reports that AI systems are increasingly automating their own development, with AI now completing tasks that previously took humans hours or days—and this capability is accelerating at an unprecedented rate. Anthropic engineers now ship 8x more code per quarter than in 2021-2025, with AI agents handling autonomous coding, testing, and task delegation. While recursive self-improvement—where AI systems could fully design their own successors—could accelerate technological progress in science and healthcare, it also raises significant risks around human control and oversight of increasingly autonomous AI systems.