Agent S4 evolves interfaces and leads open-source OSWorld V2
Original title: Evolving Agent Interfaces
Qwen3.8-27B with evolved interfaces scores 21.3%, above GPT-5.5 and Claude Opus 4.8 at far lower cost.
Agent S4 builds Evolving Agent Interfaces from unlabeled task trajectories and selects repositories by performance on unseen validation tasks. On OSWorld V2, Qwen3.8-27B with the evolved repository reaches 21.3% binary score, 3.8 times the same model with fixed interfaces, above GPT-5.5 at 13.0% and Claude Opus 4.8 at 20.6%. On ScienceBoard the same model rises from 41.4% to 58.6%. The code is open-sourced.
- Reported cost is about 4.3% of GPT-5.5 and 1.5% of Claude Opus 4.8.
- The meta-agent revises interfaces without success labels.
- Selection favors changes that generalize beyond seen training tasks.
- ScienceBoard success rose from 41.4% to 58.6% with one model.
- Products
- Agent S4OSWorld V2ScienceBoard
- AI models
- Qwen3.8-27BGPT-5.5Claude Opus 4.8
From the original article
1 min read at huggingface.co
Computer-use agents are limited by the interfaces humans give them. An agent may type a spreadsheet cell by cell when a single line of code would do, or print all two thousand lines of a file through a shell to find one function. Both the interfaces and the knowledge of when to use each are built by hand, which cannot keep pace with new applications, workflows, and models.
By Saaket Agashe, Yuzhe Yang, Gonzalo Gonzalez-Pumariega · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.