Hugging Face's Harness Adapters: RL-trained LoRA boosts Qwen3.8-27B on Pi agent
The Hugging Face team (Lysandre, Aritra, Lewis Tunstall's group, and others) released the Harness Adapters project: instead of debating which "model + tool framework" combo works best, train a dedicated LoRA adapter for each agent harness and swap them at runtime—adapting the model to the harness rather than the other way around. The first adapter (Qwen3.8-27B-pi-adapted), built on Qwen3.8-27B for the Pi agent, shows striking results.
Confirmed
- Training approach: keep the agent harness unchanged and adapt the model instead, training a small LoRA with safe RL inside the harness actually used in production
- Training recipe is public: every RL rollout is a real Pi session in the HF sandbox; the trainer does not re-implement the agent loop internally, using TRL's AsyncGRPOTrainer and OpenEn… (details compiled and published by author yb2698)
- Benchmarks: SWE-bench Pro solve rate jumps from 29.4% to 62.7%, with 55% fewer tool calls
- Case comparison: on a real SWE-Bench Pro task, the base model exhausted the 100-step limit without finishing, while the adapted version solved it in just 8 tool calls
- Deployment: the LoRA runs on transformers serve, vLLM, sglang, llama.cpp, MLX on macOS, or any other platform, and can also be merged into the base model; full details are public
Why it matters
- The method turns "model–harness fit" into a trainable, reproducible engineering problem, rather than relying on prompt tweaking or framework debates
- Especially valuable for resource-constrained settings: one base model can load different adapters on demand for different harnesses
- Publishing the full training recipe and deployment details lets the community reproduce the approach and extend it to other harnesses
2026-11-17 ~ 2026-11-18 · 8 related posts
- Episode 1: Hugging Face's Harness Adapters: Cheap LoRA+RL Tune Small Models per Agent Framework(2026-11-06, 18 posts)
- Episode 2: Hugging Face's Harness Adapters: RL-trained LoRA boosts Qwen3.8-27B on Pi agent(2026-11-17, 8 posts)
Primary sources
- Harness Adapters: RL-trained LoRA lifts Qwen3.8-27B on Pi from 29% to 63% on SWE-bench Pro — LysandreJik ·
- Pi harness adapter via RL: SWE-bench Pro 29.4%→62.7% with 55% fewer tool calls — yb2698 ·
- harness-hub releases first RL-trained harness adapter for agentic coding, full recipe coming Friday — yb2698 ·
- Real SWE-Bench Pro task: base model hits 100-step limit, adapter solves it in 8 calls — yb2698 · 2026-11-17
- [source] Pi harness adapter via RL: SWE-bench Pro 29.4%→62.7% with 55% fewer tool calls — yb2698 · 2026-11-17
- Pi harness adapter ships as a LoRA runnable on vLLM, llama.cpp, and MLX — yb2698 · 2026-11-17
- [source] harness-hub releases first RL-trained harness adapter for agentic coding, full recipe coming Friday — yb2698 · 2026-11-17
- [source] Harness Adapters: RL-trained LoRA lifts Qwen3.8-27B on Pi from 29% to 63% on SWE-bench Pro — LysandreJik · 2026-11-17
- LoRA adapters tuned per harness double Qwen3.8-27B agent benchmark scores — yb2698 · 2026-11-17
- Hugging Face ships a Pi-agent LoRA for Qwen3.8-27B, cutting tool-call flailing — Disastrous-Work-1632 · 2026-11-17
- HF CEO: RL-Adapted Qwen3.8-27B Doubles SWE-bench Pro Score to 62.7% for pi Agent — victormustar · 2026-11-18