Content Deep Dive

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

Blog post from Hugging Face

Post Details
Company
Date Published
Author
Younes B, ben burtenshaw, Nathan Habib, Lysandre, Amine Dirhoussi, Adithya S K, shaun smith, Sergio Paniego, Joel Niklaus, and Lewis Tunstall
Word Count
4,639
Company Posts That Month
68
Language
-
Hacker News Points
-
Post removed?
No
Summary

Hugging Face researchers trained reinforcement-learning LoRA adapters for Qwen3.8-27B inside three agent harnesses—Pi, Hermes Agent, and OpenCode—so training reflected each harness’s actual tools and behavior. They assembled and validated coding tasks from existing datasets and generated tasks, filtering for examples the base model solved inconsistently, then trained and evaluated the adapters on Terminal-Bench 2.1 and the hardest SWE-bench Pro tasks. Each specialist adapter improved pass rates over the base model in its own harness while reducing tool calls by roughly 53–72% on tasks both models solved; the researchers also merged the three adapters into a single “super-adapter” that improved results across all three harnesses, though specialists retained advantages in some settings. The adapters showed limited change on a separate business-automation benchmark, suggesting that gains did not clearly generalize to unrelated tasks, and trajectory analysis raised a question about whether non-thinking models learned to emit text resembling reasoning. The team released the adapters and supporting training artifacts and presented the recipe as a reproducible approach for adapting models to specific tasks and harnesses.

Trends Found in this Post
Trend Post Mentions Total Month Mentions Posts Companies MoM
AI Model Fine-tuning 8 645 141 60 -26%
LLM 4 3,674 823 199 -44%
Data Pipeline 1 176 82 43 -57%
Real-time 1 3,194 698 199 -24%
Use This Data

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.