Releases·Tool Release
The ultimate guide to multi-harness RL
Lewis from Hugging Face's post-training team released a long-form guide on training open models to perform across different coding harnesses. The guide uses open-source libraries TRL and the Harbor framework for RL environments, and offers a recipe for optimizing any open model on custom harnesses like Pi + extensions.
78 engagement1 sourceSat, Oct 3, 2026, 10:39 AM
AI Overview
- Hugging Face's post-training team (Lewis) published a guide on training open models across coding harnesses using TRL and the Harbor RL framework
- The guide targets custom harnesses like Pi + extensions, offering recipes to maximize performance with open models
- Guide is hosted at huggingface.co/spaces/FineEnvs/multi-harness-rl
Related
Developments·Paper
OpenForgeRL: Open-Source Framework for Training Harness-Based Agents End-to-End
Researchers introduced OpenForgeRL, an open-source framework that enables end-to-end training of AI agents using complex inference harnesses like Claude Code and Codex. It uses a lightweight proxy to record model calls during harness inference, allowing standard SFT/RL stacks to train stateful, multi-process agents. This addresses a key gap in open infrastructure for training modern tool-using agents.
Researchers introduced OpenForgeRL, an open-source framework that enables end-to-end training of AI agents using complex inference harnesses like Claude Code and Codex. It uses a lightweight proxy to record model calls during harness inference, allowing standard SFT/RL stacks to train stateful, multi-process agents. This addresses a key gap in open infrastructure for training modern tool-using agents.
0 engagement1 sourcearxivThu, Jul 23, 2026, 05:38 PM
Developments·Paper
Agent Lightning v1.0: Towards Harnessed Agentic RL
Agent Lightning v1.0 introduces harnessed agentic RL, where the deploy-time harness directly participates in model post-training. The original Agent Lightning's disaggregated architecture, connecting arbitrary agents to RL training via an LLM endpoint proxy, has been adopted by frameworks like verl Uni-Agent, AReaL 2.0, slime, and Polar.
Agent Lightning v1.0 introduces harnessed agentic RL, where the deploy-time harness directly participates in model post-training. The original Agent Lightning's disaggregated architecture, connecting arbitrary agents to RL training via an LLM endpoint proxy, has been adopted by frameworks like verl Uni-Agent, AReaL 2.0, slime, and Polar.
0 engagement1 sourcearxivTue, Aug 18, 2026, 08:50 AM
Releases·Tool Release
Ant Group engineers release AReno, a single-node RL post-training toolkit
Engineers from Ant Group's ASystem Team released AReno, an open-source toolkit for RL post-training of LLMs on a single node. It supports RL, SFT/DPO-style training, serving, and agentic RL without external backends, aiming to make LLM RL accessible to researchers and developers.
Engineers from Ant Group's ASystem Team released AReno, an open-source toolkit for RL post-training of LLMs on a single node. It supports RL, SFT/DPO-style training, serving, and agentic RL without external backends, aiming to make LLM RL accessible to researchers and developers.
1 engagement1 sourceredditTue, Aug 18, 2026, 05:49 AM
Developments·Paper
Applied Compute's Raymond Feng details post-training RL for enterprise models
Raymond Feng of Applied Compute describes a post-training approach where custom models keep learning on the job via reinforcement learning. An orchestrator routes interactions to inference engines, collects graded rollouts, and feeds a training engine that updates weights using a GRPO-style loop. This targets real multi-turn, long-horizon tasks in enterprise harnesses.
Raymond Feng of Applied Compute describes a post-training approach where custom models keep learning on the job via reinforcement learning. An orchestrator routes interactions to inference engines, collects graded rollouts, and feeds a training engine that updates weights using a GRPO-style loop. This targets real multi-turn, long-horizon tasks in enterprise harnesses.
18 engagement1 sourceyoutubeFri, Jul 31, 2026, 10:30 PM
Products·launched
Train a harness around a frozen LLM to improve task performance
This project introduces a method to keep an LLM frozen and train a 'harness' — a Python file that manages prompts, context, tools, and repair loops. The harness evolves via agent-proposed diffs evaluated on a task panel, using Pareto optimization. It aims to improve LLM performance on benchmarks like SWE-Bench and Terminal Bench without modifying the model itself.
This project introduces a method to keep an LLM frozen and train a 'harness' — a Python file that manages prompts, context, tools, and repair loops. The harness evolves via agent-proposed diffs evaluated on a task panel, using Pareto optimization. It aims to improve LLM performance on benchmarks like SWE-Bench and Terminal Bench without modifying the model itself.
1 engagement1 sourcehackernewsTue, Jul 21, 2026, 04:41 PM