Human-Agent-Society/reef
Reef packages the loop for agents to serve traffic, collect feedback, train or revise themselves, and ship versioned updates.
The repo describes two update surfaces: model weights, using stacks such as Slime and SGLang, and agent harnesses such as prompts, rules, and skills. Its cycle is serve, observe, grow, and commit, with feedback matched to recorded interactions before a candidate update is evaluated and published. The examples show OpenAI- and Anthropic-compatible inference endpoints, feedback reports tied to response receipts, and live updates without restarting the serving runtime. GitHub · LLM repos' note
The repo describes two update surfaces: model weights, using stacks such as Slime and SGLang, and agent harnesses such as prompts, rules, and skills. Its cycle is serve, observe, grow, and commit, with feedback matched to recorded interactions before a candidate update is evaluated and published. The examples show OpenAI- and Anthropic-compatible inference endpoints, feedback reports tied to response receipts, and live updates without restarting the serving runtime. GitHub · LLM repos' note
score 4