Reef is a continual learning infrastructure for self-improving agents that connects agent inference, feedback, learning, and versioned delivery.
Reef solves the problem of keeping deployed agents improving over time by automating the cycle of serving requests, collecting feedback, training updates, and managing versions. It processes each learning cycle in four steps: serving live traffic, recording feedback from interactions, producing model or harness updates, and committing accepted changes to version history. The tool supports multiple learning paths depending on your goal—model weight training using frameworks like Slime and SGLang, harness optimization for improving prompts and rules without local GPUs, or test-time training for scientific discovery tasks.
Reef suits teams building agents that need to evolve beyond their initial weights. Choose it if you want your agent to improve continuously from user interactions and have either a trainable model with GPU support for weight training, or a model endpoint with representative tasks and an evaluator for harness optimization. Unlike inference engines such as vLLM or SGLang that only serve traffic, and unlike RL training frameworks that only train weights, Reef uniquely combines live serving, weight training, version management, and the ability to evolve agent components beyond weights—including skills and harness elements.
The project maintains active development with regular updates to its infrastructure and documentation. The team actively engages with the community through multiple channels and maintains a public roadmap for planned features.
