Agent S4 evolves computer-use interfaces

The paper’s key insight is that interface design—not just model scale or agent prompting—can become a learned capability.

• –A meta-agent builds and revises interface tools and usage knowledge without success labels.
• –Population-based evolution tests repositories on unseen validation tasks to favor generalization.
• –The system outperformed reported GPT-5.5 and Claude Opus 4.8 scores at substantially lower cost.
• –ScienceBoard performance improved from 41.4% to 58.6% with the same model.
• –Results remain benchmark-dependent, and a 21.3% OSWorld score shows computer use is still far from reliable.