Reef’s builders say it keeps agent requests running during learning

Reef’s builders say their infrastructure separates serving agent requests from ongoing learning. Each request stays tied to its starting version, even if a new one is published before it finishes. Both model weights and the agent harness—its prompts, rules, skills, tools and orchestration—can evolve through background work. The team says a current request never waits for training, harness evolution or checkpoint export; a later request sees the new version.