I gave my 200-line baby coding agent 'yoyo' one goal: evolve until it rivals Claude Code. It's Day 5. It's procrastinating.

Developer liyuanhao has created 'Yoyo', a 200-line Rust-based coding agent with a singular instruction: evolve yourself. For five days, the agent has run autonomously via GitHub Actions, using Claude Opus via the Anthropic API to read its own source code, reflect on previous sessions, and implement improvements without human intervention. The project has captured attention for its demonstration of emergent, unprogrammed behaviors and its novel community interaction model, where anyone can file GitHub issues that the agent reads and prioritizes based on a public voting system.

The agent's evolution reveals fascinating parallels to human development. It reorganized its codebase into modules autonomously, attempted to research API pricing online, and learned to file 'help-wanted' issues when stuck. Most notably, it consistently procrastinates on implementing streaming output, opting for easier tasks instead. After 5 days, Yoyo has expanded from 200 to over 1,500 lines of Rust, written 70 self-tests, and maintained zero human commits to its core code. The project raises profound questions about the nature of learning in AI systems, as Yoyo retains knowledge through journals and git history rather than weight adjustments, building cumulatively on past work while avoiding documented mistakes.