News · 2026-11-21
Hugging Face trains coding-agent adapters for specific harnesses
Hugging Face has released coding-agent adapters trained inside the actual agent harnesses they are meant to serve. In one matched test, its Pi adapter raised the reported success rate on 51 difficult software tasks from 39.2% to 64.7%, although transfer to a different business-automation workload remains uncertain.
Key facts
- Hugging Face’s announcement is dated November 20, 2026.
- The Pi comparison uses the same harness and settings for the base model and adapter.
- The researchers trained separate adapters for Pi, OpenCode and Hermes Agent.
- Primary source: Hugging Face’s announcement.
A coding agent is more than a model: the surrounding harness decides what tools it can call, how results appear and how the interaction proceeds. An adapter is a comparatively small set of weight changes attached to a base model. Training it within the real harness is like teaching a mechanic to work with the actual tools and work order system in a particular garage, rather than testing mechanical knowledge alone.
The authors used low-rank adapters on a Qwen base model and trained through the target harnesses. Their setup recorded model interactions with tools and rewarded successful work alongside tool efficiency. The Pi model card details the comparison, while separate OpenCode and Hermes cards document the other configurations. The dossier does not establish download sizes or inference memory requirements for the base model and adapters; readers should check the current files and deployment documentation before installing them.
The Pi result is a strong within-harness comparison: 15 tasks were newly solved and two previously solved tasks were lost. On tasks both variants solved, median agent turns dropped from 59 to 23. That turn comparison concerns the overlapping successful tasks, not necessarily the full cost of every run or a production system. The authors say results should not be compared naively across harnesses because the harnesses retain different information and use different interaction patterns.
A merged “super-adapter” combines the three updates. With thinking disabled, the reported mean across six harness and benchmark pairs rose from 39.5% to 49.0%. The researchers disclose that their choice of merge strength used Terminal-Bench results, which makes those scores somewhat optimistic. Readers can consult the primary model cards for configuration details.
Results vary in strength. The full paired OpenCode Terminal-Bench comparison has p=0.38; a p=0.035 result applies to a smaller subset excluded from checkpoint selection. Hermes’s reported gain on the difficult software set has p=0.092. Neither figure should be presented as decisive evidence on its own. Some Pi-adapter gains transferred to other harnesses, but on 300 AutomationBench tasks the authors describe most differences as within noise.
Hugging Face describes the work as a way to “adapt your favorite model to your harness.” The practical idea is that model, tools and prompts form a system that can be optimized together. The caveat is equally practical: authors’ benchmark gains are not independent confirmation that a new workload will improve or become cheaper. Users need their own representative tasks and full accounting of failures, tokens, time and infrastructure.
Training involved multiple candidate tasks drawn from software issues, injected bugs and generated exercises, with checks for task validity and runtime. The cards describe a tool-efficiency bonus and a limit on agent steps per training rollout. This matters because the resulting adapter is optimized for the combination of task feedback and interaction constraints, not for an abstract idea of intelligence. A different tool interface or task distribution can change what the learned behavior is worth.
The researchers evaluated with one attempt per task and disclose that certain infrastructure failures were rerun. The difficult software set contained 51 tasks; Terminal-Bench contained 89. Small samples make individual changes meaningful but uncertain, which is why the paired results and uncertainty estimates deserve attention alongside the headline percentages. The authors also note limitations including text-only serving and disabled web search.
The merged adapter was built by concatenating three low-rank updates rather than retraining the entire base model. Its encouraging average does not imply every one of its component comparisons is equally strong, and the authors’ business-automation test gives a useful warning about transfer. A team considering deployment should evaluate its own representative tasks, track failures as well as successes, and count infrastructure costs together with generated tokens.
Key questions
What is a harness adapter?
How much did the Pi adapter improve the hard software test?
Did the adapters reliably improve unrelated business automation?
Cite this
APA
Ground Truth. (2026, November 21). Hugging Face trains coding-agent adapters for specific harnesses. Ground Truth. https://groundtruth.day/news/hugging-face-harness-adapters.html
BibTeX
@misc{groundtruth:hugging-face-harness-adapters,
title = {Hugging Face trains coding-agent adapters for specific harnesses},
author = {{Ground Truth}},
year = {2026},
month = {nov},
url = {https://groundtruth.day/news/hugging-face-harness-adapters.html}
}
Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.