OpenForgeRL: Train Harness-native Agents in Any Environment

OpenForgeRL is an open-source end-to-end training framework for harness-based AI agents.

It uses a proxy to capture model calls from agent harnesses and a Kubernetes rollout system to train across diverse environments.

Researchers reported strong results on tool-use and GUI agent benchmarks with relatively small training budgets.

By training in the same real harnesses and environments used at deployment, it could improve agent reliability and scalability.