Reef Open-Source Tool Unifies AI Agent Learning and Versioned Deployment

Human-Agent-Society has released Reef, an open-source infrastructure platform that integrates AI agent inference, feedback collection, learning, and versioned delivery into a single system. The tool organizes its workflow into four stages — Serve, Observe, Grow, and Commit — allowing developers to manage both model weight updates and agent harness changes, such as prompts and skills, within one pipeline. Reef supports two update paths: weight-oriented training using GPU stacks via Slime and SGLang, and harness-oriented updates that rely on a model API endpoint without requiring local GPUs. A key design feature is that completed training or edits do not automatically go live; candidate updates must pass a configured evaluation and selection policy at the Commit stage before being published. The project notes that updates can be applied without restarting Reef, though the effectiveness of its release controls depends on the quality of tasks and evaluators that operators configure.