DaizeDong Publishes self-evolve Harness for AI Agents to Improve Repositories Under External Gates

A public GitHub project called self-evolve packages a way for AI agents to iteratively modify skills, repositories and workflows while a deterministic harness decides whether a change is actually accepted. The repo says its core contract is that an LLM can propose patches, but accept, reject, rollback and signal selection are handled by code using external evaluation evidence, not by the model grading itself.