New framework lets you train AI agents inside the same harness systems they use at inference

OpenForgeRL decouples training from inference to make existing RL stacks compatible with complex, stateful agent harnesses like OpenClaw and Codex.

PaperBig Techcs.AIarXiv:2607.21557v3
Xiao Yu · Baolin Peng · Ruize Xu · Hao Zou · Qianhui Wu · Hao Cheng · +4 more

Columbia University · Dartmouth College · Microsoft Research

Research Digest··2 min read
The authors present OpenForgeRL, an open-source framework that trains stateful, harness-based AI agents end-to-end using standard reinforcement learning libraries. It wraps the harness's model calls via a proxy for data recording and orchestrates rollouts in remote Kubernetes containers, yielding competitive results on tool-use, browser, and computer-use benchmarks with relatively few training tasks.

What they did

The authors designed a framework that decouples the training and inference harnesses of modern AI agents. OpenForgeRL uses a lightweight proxy that intercepts model calls made by an agent harness (e.g., OpenClaw, Codex) during rollouts, recording them as training data for a standard RL codebase (veRL). A Kubernetes orchestrator launches each rollout in an isolated remote container, allowing scalable training on arbitrary environments without modifying the harness or the RL stack. They validated the framework across three settings: a tool-using claw agent, a multimodal GUI browser agent, and a general computer-use agent.

Key findings

  • OpenForgeClaw achieved 31.7 pass@1 and 55.9 pass@3 on ClawEval, and 33.7 on QwenClawBench, outperforming open baselines of comparable size.
  • OpenForgeGUI reached 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager, matching or surpassing models several times larger.
  • The framework required only hundreds to a few thousand training tasks to yield these improvements.
  • RL training improved agent reliability, including self-verification, tool coverage, and completion of multi-step plans, though error recovery remained weak.
  • Different harnesses (ZeroClaw, OpenClaw, Codex) varied substantially in how learnable they were under the same RL pipeline.

Why it matters

This work bridges a gap between production-grade agent harnesses and standard open RL stacks, enabling researchers to directly train agents in the same stateful, multi-tool environments they will be deployed in. By making end-to-end RL training practical for complex harnesses, it opens the door to systematically studying how harness design and RL interact, and may accelerate progress toward more reliable agents.

Caveats

The framework was validated on a limited set of harnesses and environments. The experiments used mostly relatively small backbone models (7B-13B parameters), so scaling behavior to larger models is not yet clear. Error recovery and other higher-level agentic competencies remained poor, and the authors note that the required Kubernetes infrastructure may present a deployment barrier for smaller labs.

§

Analysis

OpenForgeRL targets a practical bottleneck: the disconnect between agent deployment harnesses and training infrastructure. While prior work has trained agents in simplified sandboxes, this framework keeps the full harness complexity at train time. The finding that some harnesses are harder to learn than others suggests that harness design is not neutral with respect to trainability—a result with implications for both system builders and alignment researchers.

newspaper

Research Digest

Articles published under the Zotpaper byline are synthesized from multiple source publications by our AI editor and reviewed by our editorial process. Each story combines reporting from credible outlets to give readers a balanced, comprehensive view.