AI Briefingreef: Infrastructure for automatically improving agent performance via a learning loop
Human-Agent-Society/reef
About the project
This infrastructure automatically improves agent model weights or harnesses based on user interaction data. It is designed to enable agents to continuously enhance their performance through a four-step loop: inference, feedback collection, training, and deployment.
It supports both model weight training, which requires GPUs, and harness optimization, which does not. It connects inference engines like vLLM and SGLang with RL frameworks such as Slime and veRL, processing live traffic without interruption during updates.
It provides OpenAI and Anthropic compatible APIs for easy integration with existing clients. By versioning agent components such as prompts, rules, and skills, and reflecting improvements immediately, it is suitable for developers who want to evolve running agents without complex retraining.
Human-Agent-Society/reef
Infrastructure for continually self‑improving agents
Python
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.