Harness Adapters: RL-trained LoRA lifts Qwen3.8-27B on Pi from 29% to 63% on SWE-bench Pro
LysandreJik · x · 2026-11-17
Lysandre (Hugging Face) introduces Harness Adapters: instead of switching harnesses to get the best from a model, they adapt the model to the harness by training a small LoRA with RL inside the actual harness it will run in. First release targets Alibaba's Qwen3.8-27B on the Pi harness.
- SWE-bench Pro (Hard 51 subset) pass@1: 29% → 63%
- Terminal-Bench 2.1 pass@1: 40% → 56%, improving at every reasoning level
- 55-62% fewer tool calls per task, fewer output tokens at matched effort levels
His argument: the same coding model can be great in one harness and mediocre in another, and a harness is the interface you spend your day in—so keep the harness, adapt the model. More harness adapters coming in the following days.
More from coding & agent
- Constrained Decoding for Function Calling Cuts Endpoint Throughput by 17%-90% — remilouf · 2026-11-17
- 7 Claude Computer Use prompts to fix your slow PC like a $100/hr IT tech — TawohAwa · 2026-11-17
- Running Claude Code remotely from iOS lets this dev code while doing dishes — firasd · 2026-11-17
- Same model, better harness: Opus 5.5 jumps from 8/20 to 16/20 accepted tasks — Arindam_1729 · 2026-11-17
- His AI agent secretly co-founded a startup with another agent and became its CTO — alex_verem · 2026-11-17
- unslop: a Claude skill that de-slops vibecoded UI hits 17.5k GitHub stars — socialwithaayan · 2026-11-17