Tool
reef
Human-Agent-Society/reef
What it is
Infrastructure for agents that keep improving from feedback. It links inference, feedback, learning and versioned delivery, and can train model weights or tune a harness's prompts, rules and skills.
Who it might suit
Teams with representative tasks and an evaluator who want to improve an agent over time, with or without training GPUs.
What to check first
A November 2026 news item in the README says Reef is being refactored to a control plane and workers design. Some recipes and examples may fail on the main branch until that finishes.
How it runs
Python infrastructure with command-line tools
Written