Introducing harness adapters: adapt your favorite model to your
Back to Articles a]:hidden"> Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost Published November 20,…
Summary
Back to Articles a]:hidden"> Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost Published November 20, 2026 Update on GitHub Upvote 8 +2 Younes B ybelkada Follow ben burtenshaw burtenshaw Follow Nathan Habib SaylorTwift Follow Lysandre lysandre Follow Amine Dirhoussi aminediroHF Follow Adithya S K AdithyaSK Follow shaun smith evalstate Follow Sergio Paniego sergiopaniego Follow Joel Niklaus joelniklaus Follow Lewis Tunstall lewtun Follow:last-child]:mb-0"> openenv validate</code>","children":[],"isValid":true,"title":"Validation with <code>openenv validate</code>"},{"id":"difficulty-check-with-the-base-model","label":"Difficulty check with the base model","children":[],"isValid":true,"title":"Difficulty check with the base model"}],"isValid":true,"title":"Generating custom RL tasks on Hugging Face"},{"id":"testing-the-filters-with-short-training-runs","label":"Testing the filters with short training runs","children":[{"id":"the-dataset","label":"The dataset","children":[],"isValid":true,"title":"The dataset"}],"isValid":true,"title":"Testing the filters with short training runs"},{"id":"asyncgrpotrainer-for-more-efficient-grpo-training","label":"<code>AsyncGRPOTrainer</code> for more efficient GRPO training","children":[],"isValid":true,"title":"<code>AsyncGRPOTrainer</code> for more efficient GRPO training"},{"id":"training-setup","label":"Training setup","children":[{"id":"training-infrastructure","label":"Training infrastructure","children":[],"isValid":true,"title":"Training infrastructure"},{"id":"final-mixture","label":"Final mixture","children":[],"isValid":true,"title":"Final mixture"},{"id":"safeagents-in-the-loop","label":"safeagents in the loop","children":[],"isValid":true,"title":"safeagents in the loop"}],"isValid":true,"title":"Training setup"},{"id":"super-adapter-merging-all-adapters-into-one","label":"Super-adapter: merging all adapters into one","children":[{"id":"adapter-merging-evaluation-results","label":"Adapter merging evaluation results","children":[],"isValid":true,"title":"Adapter merging evaluation results"}],"isValid":true,"title":"Super-adapter: merging all adapters into one"},{"id":"final-results","label":"Final results","children":[{"id":"training-results","label":"Training results","children":[],"isValid":true,"title":"Training results"},{"id":"evaluation-results","label":"Evaluation results","children":[],"isValid":true,"title":"Evaluation results"}],"isValid":true,"title":"Final results"},{"id":"open-question-what-did-the-model-really-learn-to-do","label":"Open question: What did the model really learn to do? ","children":[],"isValid":true,"title":"Open question: What did the model really learn to do? "},{"id":"closing-words-and-future-directions","label":"Closing words and future directions","children":[],"isValid":true,"title":"Closing words and future directions"}],"classNames":"not-prose hidden lg:block"}"> Combining RL on a custom task distribution with your favorite agent harness yields more accurate and efficient models for that target harness and task distribution. Models are not fully optimized for all harnesses and the tasks you’re doing.
Positives
- Back to Articles a]:hidden"> Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost Published November 20, 2026 Update on GitHub Upvote 8 +2 Younes B ybelkada Follow ben burtenshaw burtenshaw Follow Nathan Habib SaylorTwift Follow Lysandre lysandre Follow Amine Dirhoussi aminediroHF Follow Adithya S K AdithyaSK Follow shaun smith evalstate Follow Sergio Paniego sergiopaniego Follow Joel Niklaus joelniklaus Follow Lewis Tunstall lewtun Follow:last-child]:mb-0"> openenv validate</code>","children":[],"isValid":true,"title":"Validation with <code>openenv validate</code>"},{"id":"difficulty-check-with-the-base-model","label":"Difficulty check with the base model","children":[],"isValid":true,"title":"Difficulty check with the base model"}],"isValid":true,"title":"Generating custom RL tasks on Hugging Face"},{"id":"testing-the-filters-with-short-training-runs","label":"Testing the filters with short training runs","children":[{"id":"the-dataset","label":"The dataset","children":[],"isValid":true,"title":"The dataset"}],"isValid":true,"title":"Testing the filters with short training runs"},{"id":"asyncgrpotrainer-for-more-efficient-grpo-training","label":"<code>AsyncGRPOTrainer</code> for more efficient GRPO training","children":[],"isValid":true,"title":"<code>AsyncGRPOTrainer</code> for more efficient GRPO training"},{"id":"training-setup","label":"Training setup","children":[{"id":"training-infrastructure","label":"Training infrastructure","children":[],"isValid":true,"title":"Training infrastructure"},{"id":"final-mixture","label":"Final mixture","children":[],"isValid":true,"title":"Final mixture"},{"id":"safeagents-in-the-loop","label":"safeagents in the loop","children":[],"isValid":true,"title":"safeagents in the loop"}],"isValid":true,"title":"Training setup"},{"id":"super-adapter-merging-all-adapters-into-one","label":"Super-adapter: merging all adapters into one","children":[{"id":"adapter-merging-evaluation-results","label":"Adapter merging evaluation results","children":[],"isValid":true,"title":"Adapter merging evaluation results"}],"isValid":true,"title":"Super-adapter: merging all adapters into one"},{"id":"final-results","label":"Final results","children":[{"id":"training-results","label":"Training results","children":[],"isValid":true,"title":"Training results"},{"id":"evaluation-results","label":"Evaluation results","children":[],"isValid":true,"title":"Evaluation results"}],"isValid":true,"title":"Final results"},{"id":"open-question-what-did-the-model-really-learn-to-do","label":"Open question: What did the model really learn to do?
- ","children":[],"isValid":true,"title":"Open question: What did the model really learn to do?
Risks & concerns
- Back to Articles a]:hidden"> Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost Published November 20, 2026 Update on GitHub Upvote 8 +2 Younes B ybelkada Follow ben burtenshaw burtenshaw Follow Nathan Habib SaylorTwift Follow Lysandre lysandre Follow Amine Dirhoussi aminediroHF Follow Adithya S K AdithyaSK Follow shaun smith evalstate Follow Sergio Paniego sergiopaniego Follow Joel Niklaus joelniklaus Follow Lewis Tunstall lewtun Follow:last-child]:mb-0"> openenv validate</code>","children":[],"isValid":true,"title":"Validation with <code>openenv validate</code>"},{"id":"difficulty-check-with-the-base-model","label":"Difficulty check with the base model","children":[],"isValid":true,"title":"Difficulty check with the base model"}],"isValid":true,"title":"Generating custom RL tasks on Hugging Face"},{"id":"testing-the-filters-with-short-training-runs","label":"Testing the filters with short training runs","children":[{"id":"the-dataset","label":"The dataset","children":[],"isValid":true,"title":"The dataset"}],"isValid":true,"title":"Testing the filters with short training runs"},{"id":"asyncgrpotrainer-for-more-efficient-grpo-training","label":"<code>AsyncGRPOTrainer</code> for more efficient GRPO training","children":[],"isValid":true,"title":"<code>AsyncGRPOTrainer</code> for more efficient GRPO training"},{"id":"training-setup","label":"Training setup","children":[{"id":"training-infrastructure","label":"Training infrastructure","children":[],"isValid":true,"title":"Training infrastructure"},{"id":"final-mixture","label":"Final mixture","children":[],"isValid":true,"title":"Final mixture"},{"id":"safeagents-in-the-loop","label":"safeagents in the loop","children":[],"isValid":true,"title":"safeagents in the loop"}],"isValid":true,"title":"Training setup"},{"id":"super-adapter-merging-all-adapters-into-one","label":"Super-adapter: merging all adapters into one","children":[{"id":"adapter-merging-evaluation-results","label":"Adapter merging evaluation results","children":[],"isValid":true,"title":"Adapter merging evaluation results"}],"isValid":true,"title":"Super-adapter: merging all adapters into one"},{"id":"final-results","label":"Final results","children":[{"id":"training-results","label":"Training results","children":[],"isValid":true,"title":"Training results"},{"id":"evaluation-results","label":"Evaluation results","children":[],"isValid":true,"title":"Evaluation results"}],"isValid":true,"title":"Final results"},{"id":"open-question-what-did-the-model-really-learn-to-do","label":"Open question: What did the model really learn to do?