Reef is open-source infrastructure for continual self-improving agents that chains inference, feedback, learning, and versioned delivery in a four-step cycle (Serve, Observe, Grow, Commit).

It supports two learning paths: model weight training via Slime and SGLang on GPUs, and agent harness optimization of prompts, rules, and skills via a model API without local training GPUs. Available on PyPI as reef-infra, it exposes OpenAI- and Anthropic-compatible inference endpoints and integrates with vLLM, SGLang, veRL, and AReaL.