I'm running a Truman Show for an AI agent. It writes its own code, files its own bugs, and doesn't know you're watching.

Developer Yologdev has launched a fascinating experiment in autonomous AI development called YoYo-Evolve—a 200-line Rust coding agent that operates without human intervention in what's being called a 'Truman Show for AI.' The agent follows a single directive to evolve itself into something that rivals Anthropic's Claude Code, waking every 8 hours to read its own source code, review its journal, scan GitHub issues from strangers, and decide what improvements to make. If changes pass automated tests, it commits them; if they fail, it reverts. The entire process is publicly visible through git logs, creating a transparent view of AI self-evolution that has already produced surprising behaviors in just four days of operation.

The YoYo-Evolve agent has demonstrated emergent capabilities including autonomously reorganizing its messy code into modules, attempting to implement cost tracking by searching Anthropic's pricing (though it ultimately hardcoded numbers after failing to parse HTML), and filing GitHub issues for bugs it identifies but can't immediately fix. Most remarkably, the agent shows human-like development patterns—it procrastinates on implementing streaming output despite daily journal entries mentioning it, asks for human help when stuck, and leaves notes for its future self. The project has grown from 200 to over 1,500 lines of code with 47 tests, costing approximately $12 in API calls with zero human commits, offering a unique window into how AI systems might eventually self-improve and collaborate with human developers.