On October 21, 2026, Simular published “Agent S4: evolving the interfaces computer-use agents work through.” Agent S4 does not solve tasks itself. It studies computer-use agents and improves the interfaces they work through with population-based evolution. Its output is a repository of interfaces plus knowledge of when and how to use each one. The task agent’s model and harness stay fixed; only the repository changes.
An interface is how an agent sees a computer and acts on it—for example a GUI that takes clicks and keystrokes and returns screenshots, or a shell that takes commands and returns text. Evolution can add, revise, or remove interfaces and usage knowledge. A meta-agent reads task-agent trajectories on synthetic training tasks and rewrites the repository; held-out validation scores guide which repositories develop further.
On the full 108-task OSWorld V2 benchmark, Qwen3.8-27B with the evolved repository (no scratch VM) reaches 21.3% binary success. Simular says that is the highest binary success among open-source computer-use agents—about 3.8 times the same model with fixed interfaces (5.56%). It is also above GPT-5.5 at 13.0% and Claude Opus 4.8 at 20.6%. Simular’s table lists about $1.10 per task for the evolved no-scratch run versus $25.50 for GPT-5.5; the paper abstract puts those costs at about 4.3% and 1.5% of the two models’ reported task costs. On ScienceBoard, the paper reports success rising from about 41.4% to 58.6% with the same model under a separate evolution run.
Evolving a computer-use interface repository is not the same product path as Cursor self-hosted workers driving your desktop. For whether the OSWorld numbers reproduce on your setup, or running Agent S4 yourself, we have not tested yet. If you have, you can reach the site owner through the contact link in the footer. Related reading: Cursor Self-Hosted Workers Can Now Drive Your Desktop.