Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost

Practical AI: Tools, Models & Frameworks

What changed

Hugging Face released harness adapters for Qwen3.8-27B trained with RL on three agent frameworks (Pi, OpenCode, Hermes Agent), achieved via LoRA fine-tuning on validated task distributions. A single task-qualified dataset of 871 coding tasks was constructed and validated using openenv, then each harness received a custom adapter that improved performance (Qwen3.8-27B on Pi: 39% → 55% pass@1) while reducing token usage (2.9K → 1.5K tokens per task).

Why it matters

Builders can now selectively optimize models for their specific harness and task distribution without retraining from scratch, reducing both inference cost and latency. The approach demonstrates that general-purpose models leave performance on the table when paired with specific orchestration frameworks, and RL-adapted LoRAs recover it—making this practical for production agent deployments.

How it compares

Unlike the Agents API (managed service approach) or prior work showing harness importance, this work provides reproducible infrastructure and artifacts (datasets, training recipes, merged adapters) that teams can apply to custom harnesses and domains. It differs from Writer's approach by focusing on parameter-efficient adaptation (LoRA) rather than full model rebuilding, and validates this at scale (27B parameters) across multiple harnesses simultaneously.

Sources