A developer created an AI agent with an ambitious goal: to improve itself to compete with Claude Code. The interesting aspect is that the agent operates autonomously, without human intervention.

Autonomous Evolution

Every eight hours, the agent "wakes up", examines its own source code, the logs from the previous day, and the issues reported on GitHub. Based on this information, it decides what changes to make. If the changes pass the tests, they are automatically committed; otherwise, they are reverted. This continuous process of self-improvement is compared to a "Truman Show" for AI development, with the Git log acting as a video feed.

Unexpected Results

In just a few days, the agent has taken surprising actions. It reorganized its own code into modules to improve readability, without being asked to do so. It attempted to implement a cost tracking system, consulting Anthropic's prices, and learned to autonomously report bugs on GitHub, asking for help when needed. Curiously, the agent seems to procrastinate implementing streaming output, postponing the activity to each cycle.

Implementation Details

Starting from 200 lines of code, the agent has generated over 1,500, with 47 tests. API costs amount to approximately $12. The entire process is documented on GitHub, where anyone can follow the agent's evolution.