# OpenForgeRL: Train Harness-native Agents in Any Environment

Source: [arXiv](https://arxiv.org/abs/2607.21557v1)  
Feed7 permalink: https://feed7.dev/p/2607-21557v1-0blvz16  
Published: 2026-07-23T17:38:30.000Z  
Trust: Needs Review (needs_review)

## Why Included

OpenForgeRL trains agents inside their deployment harnesses by proxying model calls into standard RL stacks and isolating rollouts in Kubernetes containers; error recovery remains weak.

## Source Summary

OpenForgeRL places a lightweight proxy between a harness and its model calls, records those calls for a standard RL stack, and runs each rollout in a remote Kubernetes container. It supports tool agents and GUI agents while keeping inference harnesses separate from training infrastructure.

## Practical Implication

For builders, the key rethink is to train against the actual harness and environment rather than a simplified chat loop. Reported RL gains include **self-verification**, **tool coverage**, and completion of multi-step plans; harness choice also materially affects learnability.

## Agent-Ready Context

OpenForgeRL places a lightweight proxy between a harness and its model calls, records those calls for a standard RL stack, and runs each rollout in a remote Kubernetes container. It supports tool agents and GUI agents while keeping inference harnesses separate from training infrastructure.

For builders, the key rethink is to train against the actual harness and environment rather than a simplified chat loop. Reported RL gains include **self-verification**, **tool coverage**, and completion of multi-step plans; harness choice also materially affects learnability.

Results use hundreds to a few thousand tasks. OpenForgeGUI reports **37.7 on OSWorld-Verified**, **63.0 on Online-Mind2Web**, and **72.3 on WebVoyager**, but cross-benchmark values are not directly comparable, and error recovery remains weak.

## Context Map

- Layer: agent
- Domains: coding
- Topics: harness-engineering, agent-reliability, computer-use

## Uncertainty

- Results use hundreds to a few thousand tasks. OpenForgeGUI reports **37.7 on OSWorld-Verified**, **63.0 on Online-Mind2Web**, and **72.3 on WebVoyager**, but cross-benchmark values are not directly comparable, and error recovery remains weak.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.