reef
Infrastructure for continually self‑improving agents
What is reef?
Infrastructure for continually self‑improving agents
Overview of reef
What it is
Reef is an open-source infrastructure designed for continually self-improving agents. It connects agent inference, feedback, learning, and versioned delivery. Reef can train model weights or improve an agent's harness, including its prompts, rules, and skills.
Key features
- Connects agent inference, feedback, learning, and versioned delivery.
- Supports training model weights with Slime and SGLang.
- Supports evolving agent harnesses (prompts, rules, skills).
- Provides version management for artifacts.
- Stays live through updates without restart.
- Serves live traffic and integrates with inference engines like vLLM and SGLang.
- Offers OpenAI- and Anthropic-compatible endpoints.
Use cases
- When you want an agent to keep improving by learning from user interactions.
- For model weight training using a trainable model, a supported GPU stack, and feedback.
- For harness optimization using a model endpoint, representative tasks, and an evaluator without local training GPUs.
- For scientific discoveries through test-time training with an execution environment, correctness checker, and measurable objective.
Getting started
See the repository README for installation and usage instructions.
License and pricing
Licensed under Apache-2.0. Open source.
Based on the project's public README and repository metadata.