HyperAgents: Self-referential self-improving agents

Facebook Research has open-sourced HyperAgents, a self-referential agent framework where a meta-agent iteratively rewrites task-agent code to improve performance across domains. The system uses a generate loop with foundation models (OpenAI, Anthropic, Gemini) and stores experiment logs for analysis. It explicitly warns that the repo executes untrusted, model-generated code — a meaningful safety consideration for practitioners evaluating agentic coding tools.