Evo is an open-source autoresearch orchestrator for codebases. It discovers or accepts metrics, sets up evaluation, and runs coding-agent experiments in loops that keep measured improvements and discard failures.
The project adapts Karpathy-style autoresearch from machine-learning training experiments to repository work such as performance optimization, benchmark improvement, and gated code changes. Its useful constraint is that the loop depends on explicit metrics and gates; without them, agents can optimize the score while breaking the system.