When AI Builds Itself: Our Progress Toward Recursive Self-Improvement

The era of human-driven AI development is shifting. While humans have historically directed every step of the development cycle, a growing share of this work is being delegated to AI systems, a trend that is accelerating the pace of progress. Research from the Anthropic Institute indicates that this shift points toward a future of recursive self-improvement, where an AI system could autonomously design and develop its own successor.

We are not at that point yet, and such an outcome is not inevitable. However, the trajectory is already visible in the productivity of engineering teams. Anthropic engineers now ship, on average, 8x as much code per quarter as they did from 2021-2025.

The evolution of these capabilities follows a clear progression of autonomy:

• 2021–2023: Development mirrored traditional tech work, with people writing code and documentation on laptops.
• 2023–2025: The introduction of chatbots allowed for generating short code snippets and copying output into editors.
• 2025–2026: The rise of coding agents enabled the writing and editing of code and entire files.
• Today: Autonomous agents can run code and delegate hours of work to other agents.
• 20XX?: The potential for agents to build and train models themselves, allowing future versions of Claude to be continuously improved by Claude itself.

This acceleration is not merely theoretical. The rate at which AI models improve is increasing, with the length of tasks they can reliably complete on their own doubling roughly every four months, an acceleration from an earlier trend of doubling every seven months. In March 2024, Claude Opus 3 demonstrated the ability to complete software tasks.

The implications of reaching a state of recursive self-improvement are profound. On one hand, the ability for AI to build itself could drive massive advancements in science and healthcare. On the other, it introduces the risk of humans losing control over these systems. As systems gain the capacity to build their own successors, the methods we use to secure, monitor, and shape their behavior become the most critical components of the technology's development.

The central question for policymakers and builders is no longer just how to build better models, but how to maintain oversight when the development loop begins to close itself.