Hugging Face Open-Sources Harness Adapters: Per-Agent-Framework LoRA Trained with RL

The Hugging Face team (harness-hub, with contributors yb2698 and ariG23498) released and live-streamed the training of "harness adapters": instead of modifying the agent harness itself, a small LoRA adapter is trained for each "model × agent harness" combination and optimized directly with reinforcement learning in unmodified, real coding environments (NousResearch's hermes, pi, and opencode). Covered models include LiquidAI's LFM2.5-24B-A2B (MoE, 2B active parameters), LFM3-1.7B, and thewebAI's TwIL-8B. Lysandre (Liquid AI) and Aritra (HF) presented the progress from the model side and the platform side respectively, and HF published an accompanying in-depth blog post.

Confirmed

• Single-GPU LoRA fine-tuning of LFM3-1.7B for agent RL; after a 20-step smoke run, tested on 150 never-before-seen held-out tasks: the opencode harness gained 9.4 ± 6.3 points pass@1 (69%→79%)
• Compute-saving engineering: one vLLM per model, one trainer per harness, with LoRA adapters mounted on shared weights; the three trainers for LFM3-1.7B share a single vLLM, and evaluating 1,350 agent sessions cost only $5
• Reward design: reward = task validator score + 0.1 × (1 − calls/15), applied only to correct rollouts; a coefficient of 0.3 hurt accuracy, while 0.1 reduced tool calls by 11–17% with no accuracy loss
• Environment synthesis: seeded from public agentic trajectories on HF Hub, converted into tasks only when they contain verifiable units of work; four new environment types (Repair/Build/Extend/QA) were synthesized, with LFM2.5-24B self-evaluation used for difficulty stratification, addressing the bias of older environments toward data analysis
• Training uses the safeagents anti-cheating scheme, with all tool calls audited by proxies to prevent reward hacking; the environments and training pipeline were open-sourced during the livestream

Not yet confirmed

• The exploration directions (new RL tasks from trajectory construction, new harnesses, cross-domain transfer, etc., see m11) remain goals of this round's plan; livestream results are pending

Why it matters

• The method shows that small models can achieve clear gains on a specific agent framework through extremely low-cost ($5-level) targeted RL + LoRA, offering a reproducible open-source paradigm for co-optimizing models with frameworks

2026-11-06 ~ 2026-11-06 · 16 related posts

• Livestreamed RL training of LoRA harness adapters for Hermes, pi and opencode, fully open-sourced — yb2698 · 2026-11-06
• Hugging Face livestreams open RL training: trajectory-built envs, safeagents, LoRA evals — yb2698 · 2026-11-06
• Hugging Face synthesizes RL envs from open trajectories: Repair, Build, Extend, QA — yb2698 · 2026-11-06
• Synthesizing RL Envs from Open Trajectories: LFM2.5-24B Self-Scores Task Difficulty — yb2698 · 2026-11-06
• Reward Shaping Ablation: 0.1 Bonus Cuts Tool Calls 11-17% at No Accuracy Cost — yb2698 · 2026-11-06
• Evaluating 1,350 Agent Sessions on One GPU for $5 via Shared LoRA Adapters — yb2698 · 2026-11-06
• [source] 20-Step RL on LFM3-1.7B Lifts opencode Pass@1 by 9.4 Points — yb2698 · 2026-11-06
• HF live-streams agent RL: LoRA adapters trained inside real harnesses like opencode — yb2698 · 2026-11-06
• Team livestreams open RL training of LoRA harness adapters for hermes, pi and opencode — yb2698 · 2026-11-06
• HF team live-trains harness adapters: LoRA + RL to make small models better at agentic tasks — yb2698 · 2026-11-06
• Four research directions in HF's harness adapter training: trajectory tasks, new harnesses, transfer — yb2698 · 2026-11-06
• [source] HF details harness adapters: one LoRA per agent harness, every tool call vetted by safeagents — yb2698 · 2026-11-06
• [source] Harness adapters: per-agent-framework LoRAs cut tool calls, $5 for 1,350 eval sessions — LysandreJik · 2026-11-06
• Hugging Face launches Harness Adapters: per-task LoRAs for coding harnesses — Disastrous-Work-1632 · 2026-11-06
• Hugging Face releases harness adapters: safe RL inside pi, opencode and hermes — yb2698 · 2026-11-06
• Harness adapters training goes live: RL-tuned LoRAs for LFM2.5, LFM3 and Twil-9B — yb2698 · 2026-11-06