The post introduces Reef, an open-source infrastructure framework that treats an agent’s live inference server as the substrate for continuous learning rather than a terminal deployment stage. It implements a four-stage loop (serve and collect experience, evolve model checkpoints or harness logic via pluggable ‘learning recipes,’ evaluate candidates, and hot-swap approved versions without downtime), using Git-LFS-backed versioning and a training backend (Cordis) for harness evolution alongside distributed model training. The system already supports multiple recipes spanning online RL, test-time training, and harness evolution, and is maintained at github.com/Human-Agent-Society/reef.
