harness-hub trains RL LoRA adapters so open models fit each coding agent harness
yb2698 · x · 2026-11-19
- The idea: the same open model behaves differently in Pi, OpenCode, Hermes or Claude Code because each harness has its own tools, prompts and control flow. harness-hub trains small LoRA "harness adapters" — one per model-harness pair — with reinforcement learning run inside the unmodified harness; you keep your base model and load the adapter for the agent you use.
- Recipe: TRL AsyncGRPOTrainer + custom RL environments + OpenEnv + safeagents, with GGUF weights provided.
- Coverage: adapter collections already released for Qwen3.8-27B (Pi), LiquidAI LFM2.5-24B-A2B (Pi, Claude Code, Hermes, OpenCode), LFM3-1.7B and an experimental TwIL-8B build, all open on Hugging Face.
- A full blog post with the complete training recipe drops Friday, so you can train an adapter for your own harness and tasks.
Related event: harness-hub Open-Sources RL-Trained Harness Adapters(2 posts)→
More from coding & agent
- Creator Studio ships with in-app assistants, MCP support, and open Keynote format — jasonwuishere · 2026-11-19
- Deepnote launches shareable apps to fix vibe-coded dashboards stuck in single Claude Code sessions — aakashgupta · 2026-11-19
- Kong launches Volcano, a 'self-healing' cloud for agents after handling 500T AI tokens a year — EXM7777 · 2026-11-19
- Grok bot trained on Dan Koe's system turns one idea into a week of content — VibeMarketer_ · 2026-11-19
- Matt Pocock unveils VECTOR method for adopting AI on any task, seeks early feedback — mattpocockuk · 2026-11-19
- Meta's MLE agents ran 1,097 experiments on production rankers with 29.6% acceptance rate — dair_ai · 2026-11-19