harness-hub open-sources RL-trained LoRA adapters tuned per agent harness, recipe coming Friday

yb2698 · x · 2026-11-19

harness-hub released harness adapters: one small LoRA per base model and per agent harness (Pi, Claude Code, Hermes, OpenCode), based on the insight that the same model behaves differently in each harness due to differing tools, prompts and control flow.

Training recipe: every RL rollout is a real Hermes Agent session in an HF sandbox, trained with TRL's AsyncGRPOTrainer + OpenEnv, with safeagents checks on every tool call. Adapters and GGUF weights are already open for Qwen3.8-27B, LiquidAI LFM2.5-24B-A2B, LFM3-1.7B and TwIL-8B. The author says a full blog recipe for training your own harness adapters arrives Friday.

Original post →

More from coding & agent

coding & agent channel →