Evolving Agent Interfaces

TL;DR

Agent S4 evolves computer-use interfaces and usage knowledge from task trajectories, while keeping the backbone model fixed. The authors report gains over fixed interfaces; these are system results rather than a new model release. (source)

Authors & Org

Saaket Agashe, Yuzhe Yang, Gonzalo Gonzalez-Pumariega, Bhuvan Chandra Koduru, Chih-Lun Li, Jiachen Yang, Ang Li and Xin Eric Wang. The abstract record does not establish affiliations. First submitted 2026-10-20. (source)

Method

A meta-agent examines trajectories from synthetic training tasks without success labels, then builds and revises an interface-and-usage repository. Population-based evolution chooses repositories using validation tasks the meta-agent never sees, aiming to reward changes that generalize. The code repository is Agent S; code and full paper were not reviewed. (source)

Results

With Qwen 3.8 27B, Agent S4 reaches 21.3% binary score on OSWorld V2, 3.8× the same model with fixed interfaces. The abstract compares GPT-5.5 at 13.0% and Claude Opus 4.8 at 20.6%, at about 4.3% and 1.5% of their respective costs. ScienceBoard success rises from 41.4% to 58.6% with the same backbone. These are author-reported results, not independently reproduced measurements. (source)

Significance

The study makes the agent's available interfaces and knowledge of when to use them explicit optimization targets. It extends Agents (LLM Agents) and Eval Harness Configuration by holding the backbone fixed while changing its action surface. (source)

Open Questions

The abstract does not specify total interface-evolution cost, validation-set size or whether the reported cost comparisons amortize evolution. OSWorld V2 binary scores should not be combined with partial-credit or differently versioned OSWorld results. These limits constrain how the reported gains can be interpreted. (source)

Cite

• arXiv:2610.28601
• Captured abstract and metadata