Gradient open-sources a full reinforcement-learning loop for training tool-using research agents with GRPO

GitHub·medium signalXBlueskyLinkedInCopy link

RobertGolds1/Gradient (created 2026-08-23, 423 stars in about a day, Apache-2.0, Python) is a reference system for training research agents with RL, built on OpenPipe ART. It ships a Research Environment containing a reproducible company workspace of emails, contracts, policies, meeting notes and customer records with evidence deliberately distributed across them, and a Learning Loop that records complete tool-use trajectories, scores each episode on answer correctness, citation quality and tool efficiency, trains with GRPO, then evaluates the trained adapter against the base model on held-out tasks. It is one of the few public repos that makes the whole agent-RL workflow inspectable rather than shipping just a trainer.