Reef

Reef is open source infrastructure for AI agents that keep improving from live use. It connects agent inference, feedback, model training and versioned deployment, so an agent can learn from real traffic and update both its model weights and its harness, meaning the prompts, skills and surrounding code that shape its behavior. A developer starts a local stack with a reef serve command and a recipe file, sends a provider native request, and grades the receipt Reef returns. Reef supports two kinds of improvement. For model weights it uses GPU training recipes on Slime and SGLang, including approaches such as SAO, OpenClaw RL and test time training methods like TTT Discover and Guidance TTT. For the agent harness it needs no GPU and works with any model endpoint, using recipes such as Reefine, which refines a coding harness from plain language instructions, SkillClaw, which grows skills from failures, GEPA reflective prompt evolution and Meta Harness search. It is aimed at researchers and engineering teams building agents that should adapt after deployment rather than stay fixed. The project is released under the Apache 2.0 license, is at version 0.1.2 and is developed in the open on GitHub. The site lists no pricing, as the software is free to use.