Evolving Agent Interfaces

[arXiv] Saaket Agashe, Xin Eric Wang and other authors have released Agent S4 on arXiv, an autonomous computer-use agent that can build an "evolved agent interface" from task experience. The paper states that existing computer-use agents are constrained by human hand-designed interfaces. In Agent S4, a meta-agent analyzes the task agent's operation trajectories on synthetic training tasks without success labels, builds and revises interfaces and usage knowledge in a resource library, and then, through population-based evolution, selects the resource library based on performance on validation tasks the meta-agent has never seen. The authors state that on OSWorld V2, Agent S4 with Qwen3.8-27B and the evolved interface library achieved a binary score of 21.3%, a new high for open-source computer-use agents, 3.8 times that of the same model using a fixed interface, and higher than GPT-5.5 (13.0%) and Claude Opus 4.8 (20.6%), at approximately 4.3% and 1.5% of their respective costs. On ScienceBoard, the success rate of the same model rose from 41.4% to 58.6%.