Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

OpenForgeRL: Train Harness-native Agents in Any Environment

Source

arXiv

arxiv.org/abs/2607.21557v1

Read the source

Source

Plain-Language Brief

OpenForgeRL is an open-source framework for training AI agents end-to-end in various environments using complex harnesses like Claude Code, Codex, and OpenClaw. The framework enables scaling by using a lightweight proxy to serve model calls and record them as training data, and a Kubernetes orchestrator to run each rollout in its own container.

What the Document Says

  • OpenForgeRL is an open-source framework for training harness-based AI agents end-to-end in diverse environments.
  • The framework uses a lightweight proxy and Kubernetes orchestrator to enable scaling and train on any harness in any environment.
  • It allows researchers to study and improve agents directly in the real harnesses they are deployed with.
  • The text provides validation results for OpenForgeRL across various harnesses and environments, including ClawEval, QwenClawBench, OSWorld-Verified, Online-Mind2Web, and WebVoyager.
  • OpenForgeClaw and OpenForgeGUI outperform open baselines on most benchmarks and perform as well or better than models larger in size in the GUI setting.

Key Concepts Mentioned

  • AI agents
  • harnesses
  • infrastructure
  • SFT/RL stacks
  • end-to-end training
  • OpenForgeRL
  • veRL
  • Kubernetes

Claims Made by the Document

  • OpenForgeRL allows for end-to-end training of harness-based agents in diverse environments.
  • The framework decouples training and inference, enabling easy research, study, and improvement of agents.
  • OpenForgeClaw and OpenForgeGUI outperform open baselines on most benchmarks.

Evidence Used

  • The text describes the functionality and design of OpenForgeRL.
  • Validation results are provided for OpenForgeRL across various harnesses and environments.

Methods / Approach

  • Using a lightweight proxy and Kubernetes orchestrator to train AI agents end-to-end

Limitations Stated by the Authors

Some limitations might be present, but they are not explicitly stated in the provided text.

Limitations Noted by the Agent

Agent interpretation — not from the source.

The agent's performance on critical abilities such as error recovery remains weak.

Open Questions

  • What are the specific details of the experiments conducted to validate OpenForgeRL?
  • How does the agent perform in real-world scenarios beyond benchmarks?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The text provides some validation results, but additional information might be needed to fully understand the performance and capabilities of OpenForgeRL.

Raw Abstract / Excerpt

Modern AI agents rely on elaborate inference harnesses such as Claude Code, Codex, and OpenClaw to drive multi-turn reasoning, tool use, and access to external systems. While powerful, these complex harnesses also make agents hard to train end-to-end with open infrastructure, whose SFT/RL stacks cannot natively express stateful, multi-process harness inference. To address this, we present OpenForgeRL, an open-source framework for training harness-based agents end-to-end in diverse environments. OpenForgeRL achieves this with a lightweight proxy that serves the harness's model calls while recording them as training data for a standard RL codebase (e.g., veRL), and a Kubernetes orchestrator that runs each rollout in its own remote container, together enabling training on any harness in any environment at scale. By decoupling training and inference, OpenForgeRL allows researchers to easily train, study, and improve agents directly in the real harnesses and environments they are deployed with. We validate our framework across diverse, complex harnesses and environments, spanning tool/claw-based agents and multimodal GUI browser- and computer-use agents. Using only hundreds to a few thousand tasks, OpenForgeClaw reaches 31.7 pass^3 and 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench. OpenForgeGUI reaches 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager. Both outperform open baselines of similar size on nearly all benchmarks, and in the GUI setting match or surpass models several times larger. Beyond benchmarks, we analyze how harness choice (e.g., ZeroClaw, OpenClaw, Codex) and RL shape agent behavior. We find that some harnesses are substantially harder to learn than others, and that RL improves agentic reliability, such as self-verification, tool coverage, and completing multi-step plans, though critical abilities such as error recovery remain weak.