Reef
Self-hosted AI agent infrastructure that learns from feedback, trains weights with Slime and SGLang, or improves prompts and skills without local training GPUs.
Reef is self-hosted infrastructure for developers who want AI agents to improve through feedback on actual interactions. It connects inference and learning with versioned deployment, so an agent can update its model weights or its prompts, rules, and skills while continuing to serve requests. It's open source under Apache 2.0.
The two approaches have different hardware needs. Weight training uses Slime and SGLang and requires a trainable model with a supported GPU stack. Harness optimization, which changes the instructions and tools around a model, works with any model endpoint and doesn't need local training GPUs. Reef runs the service and keeps harness state locally; if you connect an external model API, inference requests go to that provider.
Reef accepts OpenAI- and Anthropic-compatible requests and associates feedback with the interactions it evaluates. That feedback can include scores, written explanations, or structured data. Candidate updates pass through evaluation and selection before Reef publishes them, and version history tracks the resulting artifacts. Updated weights reach the serving runtime without a restart.
Its recipes cover different kinds of agent improvement. Reefine refines a coding harness from plain-language requests, while SkillClaw develops skills from failures. GEPA evolves prompts through reflection, and Meta-Harness searches across full harness designs. For model training, SAO learns from a single rollout and OpenClaw-RL learns from personalized chat.
