I Built an AI Agent That Rewrites Its Own Code

The developer created a tiny version of the Darwin Gödel Machine, an AI agent that edits its own code to improve performance on specific tasks. Initially, the agent could only perform one out of eight tasks, but after rewriting its code, it successfully completed all eight tasks. This was achieved by adding skills, such as handling messy text, which enabled the agent to solve multiple tasks simultaneously. The agent's improvement was made possible by a simple test that validated whether each change improved its performance. Researchers from the original 2025 study reported similar results, with an AI coding assistant improving its tooling and solving 20% to 50% of a hard benchmark of real GitHub issues.

The development of self-improving AI agents like this one marks a shift in AI research. For years, improving AI performance relied on making models larger. However, recent studies, including the one on which this agent is based, focus on enabling AI to improve itself while running. This approach allows AI to adapt and learn without requiring extensive retraining. Other research, such as "Language Models Need Sleep" (2026), explores ways to enable AI agents to tidy up their own memory during an offline "sleep" phase. Shridhar Shah, a Senior Software Engineer at Cisco, demonstrated this concept by building a minimal AI agent that rewrites its own code.

The implications of self-improving AI agents are significant, but they also raise concerns about safety and control. The developer of this agent emphasized that the "edits" come from a fixed list of safe skills, ensuring that no dangerous code is executed. However, as AI agents become more autonomous, it is crucial to develop robust testing and validation procedures to prevent potential risks. The developer's work, available on GitHub, provides a tangible example of how AI can improve itself through self-rewriting code.