Human-Agent-Society released Reef, an open source, Apache 2.0-licensed platform it calls continual learning infrastructure, connecting an agent's inference, feedback collection, learning and versioned deployment into one loop instead of leaving improvement to a manual retraining cycle, according to the project's repository.
Reef runs a four-step cycle: serve requests, observe feedback, grow updates and commit versioned changes. It supports two kinds of learning, updating a model's actual weights through integrations with Slime and SGLang, or updating the harness around a frozen model, its prompts, rules and skills, without touching the weights at all.
The project ships OpenAI- and Anthropic-compatible API endpoints, keeps a version history of every learned artifact with policies for accepting or rejecting an update, and covers recipes for continual learning on a stream of tasks, learning from live traffic, and test-time training for research use. It had drawn more than 3,500 GitHub stars as of this week.
Most agent deployments today are static: whatever the prompt and model were at ship time is what a user gets months later, until an engineer manually intervenes. A framework that treats adjusting the harness based on what actually happened in production as a first-class, versioned operation is a real answer to a problem every team running agents at scale eventually hits.