Reef
Open-source infrastructure connecting agent inference, feedback, and learning so an agent's model weights or its prompts, rules, and skills keep improving.
Reef processes a repeating four-step learning cycle: serve agent requests and record the interaction, match feedback back to the recorded interaction and check whether it is eligible for training, produce an update from the eligible records through a configurable recipe, and apply a selection policy before publishing an accepted update, staying live through the update. It supports training model weights with Slime and SGLang, or instead improving an agent's harness, meaning its prompts, rules, and skills, without any local training GPUs.
Compared to running an inference engine such as vLLM or SGLang, or a separate RL training framework such as Slime or veRL, Reef is designed to be the layer that also does version management and stays live through updates while the other two continue serving traffic and training weights respectively.
Its artifact and checkpoint functionality requires Git LFS, it is written in Python, and distributed under the Apache-2.0 license.