OpenForgeRL: Train Harness-native Agents in Any Environment
The paper proposes an open-source way to train agents inside the same harnesses they use at inference time.
OpenForgeRL uses a proxy to capture harness model calls as training data for standard RL stacks, plus Kubernetes containers to isolate rollouts. The authors test it across claw/tool agents and multimodal GUI browser and computer-use agents. Reported results beat similar-size open baselines on nearly all listed benchmarks, with GUI agents matching or exceeding larger models in some cases. The paper also says RL improves reliability traits like self-verification and tool coverage, while error recovery remains weak. ArXiv · AI/CL/LG's note
OpenForgeRL uses a proxy to capture harness model calls as training data for standard RL stacks, plus Kubernetes containers to isolate rollouts. The authors test it across claw/tool agents and multimodal GUI browser and computer-use agents. Reported results beat similar-size open baselines on nearly all listed benchmarks, with GUI agents matching or exceeding larger models in some cases. The paper also says RL improves reliability traits like self-verification and tool coverage, while error recovery remains weak. ArXiv · AI/CL/LG's note
score 6