Reef
Self-hosted AI agent infrastructure uses feedback to train model weights or improve prompts, rules, and skills through a model endpoint without training GPUs.
Reef is self-hosted infrastructure for developers who want AI agents to learn from their interactions. It connects live requests and feedback to changes in model weights or the prompts, rules, and skills that guide an agent.
- Feedback from interactions: Use scores, written explanations, or structured feedback tied to recorded requests to guide learning, with support for OpenAI- and Anthropic-compatible clients.
- Model training: Train weights with Slime and SGLang, using recipes such as SAO for learning from a single rollout and OpenClaw-RL for personalized chat.
- Prompt and skill improvement: Reefine refines a coding harness from plain-language requests, SkillClaw develops skills from failures, GEPA evolves prompts through reflection, and Meta-Harness searches across full harness designs.
- Versioned updates: Evaluate candidate changes, publish accepted updates, and keep artifact version history while the service continues handling requests; trained weights reach the serving runtime without a restart.
Reef is open source under Apache 2.0. Weight training requires a trainable model and a supported GPU stack. Improving prompts, rules, and skills works with any model endpoint and doesn't need local training GPUs, but it does require representative tasks and an evaluator. The service and harness state stay local; when you connect an external model API, inference requests go to that provider.
Logotyper
