Gradient open-sources a full RL training loop for research agents

RobertGolds1/Gradient went up on GitHub August 23 under an Apache-2.0 license and passed 423 stars in about a day. It's built on OpenPipe ART, and it does something most agent-RL repos skip: it publishes the environment along with the trainer.

That environment is a reproducible company workspace, emails, contracts, policies, meeting notes, customer records, with the evidence an agent needs deliberately scattered across them. An agent has to use tools to hunt through that mess and pull out a correct answer with the right citations.

The Learning Loop records every tool-use trajectory the agent takes, then scores each episode on three things: whether the answer was right, whether the citations back it up, and how efficiently the agent used its tools to get there. Training runs on GRPO, and the resulting adapter gets evaluated against the base model on tasks it hasn't seen.

Most public agent-RL work stops at the trainer and leaves the benchmark and the reward function as a black box, so nobody outside the team can check whether the gains are real or just tuned to a hidden eval. Gradient's Research Environment is inspectable start to finish. Anyone can run the same messy company workspace, see what the agent actually does with its citation checks and tool budget, and compare their own fine-tune against the same held-out tasks.

What's not in the repo yet: any numbers on how much the trained adapter beats the base model by. A day-old repo with no benchmark table is normal, but it's the first thing worth checking before treating this as more than a well-built harness.