Hugging Face Open-source multi-Harness reinforcement learning solution supports integration with Agents such as Claude Code and Codex

Article

Hugging Face Let Claude Code and Codex participate directly RL Training

Insight Beating AI breaking news:Hugging Face has open-sourced a Multi-harness RL solution that can directly connect established Coding Agent such as Claude Code, Codex, and OpenCode to the reinforcement learning workflow to train open-source models connected to it, without needing to build a separate training environment for each Agent.

Large-model companies such as OpenAI have long trained models in their own Agent environments.Hugging Face has now turned similar capabilities into a general open-source tool, enabling other teams to directly use established Agents for RL.

Moreover, models do not need to be trained in only one type of Agent. A single training run can rotate tasks across Claude Code, Codex, OpenCode, and Mini-SWE-Agent, allowing the same model to adapt to different prompts, tools, and execution logic, reducing specialization in only one Harness.

They tested a small AI-parameter model from Liquid 2600000000. With the same weights and only the Agent changed, task pass rates could fall from approximately 62%to 33%. After training across all Agent environments, the average pass rate increased from 42.2%to 54.2%, and improvements across different Agents were more balanced than when training in only a single environment.

Original link https://m.theblockbeats.info/flash/370467

Key points

Hugging Face Open-sourced a Multi-harness RL solution.
This solution can connect Claude Code, Codex, OpenCode, and other Coding Agent to reinforcement learning workflows.
A single training run can rotate among Claude Code, Codex, OpenCode, and Mini-SWE-Agent。
The test used a small model from Liquid AI with 2600000000 parameters.
After changing the Agent while keeping the same weights, the task pass rate fell from approximately 62%to 33%。
After training in Agent environments, the average pass rate increased from 42.2%to 54.2%。

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This solution treats different Coding Agents as training environments, exposing the same model to different prompts, tools, and execution methods, thereby reducing the likelihood that the model adapts to only one type of Agent.

Why it matters to readers

If the test results are reproducible, the model may perform more evenly across multiple Coding Agent environments.

It shows that model training need not rely solely on one fixed coding assistant and can instead learn across multiple assistant environments.

Risks and unknowns

The information comes from media reporting; no evidence was provided Hugging Face in the form of an official announcement, code repository, or complete experimental materials.
The task set, training configuration, and specific experimental process used in the test are not described in the evidence.
The actual integration scope and availability of Agents such as Claude Code and Codex still require verification.

Hugging Face Open-source multi-Harness reinforcement learning solution supports integration with Agents such as Claude Code and Codex

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center