yoyo-evolve
An AI agent that reads, modifies, and tests its own code, helping developers explore the frontier of autonomous software development.
Tool overview
yoyo-evolve is an open-source self-evolving AI coding agent. It wakes every few hours, scans its own repo and GitHub issues, proposes changes, and runs cargo build and cargo test. Only if CI passes does it commit; otherwise it reverts and logs the failure. This demonstrates a closed-loop observe-modify-verify-evolve cycle.
Advantages include a multi-layer safety net (protected file checks, revert mechanisms) and accumulating capabilities (skill-evolve, 70+ slash commands). No human writes code – developers steer via labeled issues as “agent-input”.
Currently experimental with gaps: no tool output streaming, incomplete permission system, limited project indexing. Autonomy depends on Claude Sonnet; the 4–8h cycle is too slow for urgent fixes, and reverts may discard valuable intermediate insights.
Deploy via curl install or cargo run with ANTHROPIC_API_KEY; for full auto-evolution, fork and use GitHub Actions. Due to code generation instability and opaque decisions, it suits AI agent researchers and harness engineering learners, not yet production codebases.