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 designed to train AI agents using complex inference harnesses like Claude Code, Codex, and OpenClaw, in various environments. The framework allows end-to-end training of these agents with standard RL codebases, making it possible to study and improve them directly in the real environments they are deployed in. The authors demonstrate its effectiveness across diverse harnesses and environments, achieving strong performance on various benchmarks.

What the Document Says

  • OpenForgeRL is an open-source framework for training AI agents using complex inference harnesses.
  • It allows end-to-end training of these agents with standard RL codebases.
  • The training can be done in diverse environments.
  • A lightweight proxy serves the harness's model calls and records them as training data.
  • A Kubernetes orchestrator runs each rollout in its own remote container.
  • OpenForgeRL allows researchers to study and improve agents directly in the real harnesses they are deployed with.

Key Concepts Mentioned

  • AI agents
  • inference harnesses
  • end-to-end training
  • open infrastructure
  • SFT/RL stacks
  • standard RL codebase
  • veRL
  • Kubernetes
  • rollout
  • remote container

Claims Made by the Document

  • OpenForgeRL can train agents in diverse, complex harnesses and environments.
  • It can outperform open baselines of similar size on various benchmarks.

Evidence Used

  • Reach of 31.7 pass^3 and 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench for OpenForgeClaw.
  • Reach of 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager for OpenForgeGUI.

Methods / Approach

  • Use of a lightweight proxy and a Kubernetes orchestrator

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

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.