The Pulse
Simular’s Agent S4 Evolves Interfaces for Computer-Use Agents
Simular introduced Agent S4 on October 21, 2026, a system that evolves the interfaces computer-use agents rely on. On OSWorld V2, the company says Qwen3.8-27B with S4’s evolved interfaces scored 21.3% on the benchmark’s 108 tasks.
AI.info Team · Written with AI · fact-checked ·
Simular introduced Agent S4 on October 21, describing a system that studies computer-use agents and evolves the software interfaces they work through. The aim is to change what an agent can do with a computer, not just make it better at clicking through a fixed set of buttons. Simular says the task agent’s model and harness remain fixed while a separate meta-agent revises a repository of interfaces and instructions.
Agent S4 changes the tools, not the model
Computer-use agents commonly receive screenshots and act through mouse clicks and keystrokes. That can force an agent to edit a spreadsheet cell by cell, or open a terminal through the desktop and type commands one character at a time. Simular’s approach lets the agent gain more direct ways to work, such as editing a block of spreadsheet cells in one operation and returning the changed values.
The system runs a task agent and a meta-agent. The task agent attempts synthetic training tasks using its current repository; the meta-agent examines its action histories, without success labels, and proposes interface or instruction changes. It can revise one repository or combine elements from two, then the task agent tests the result against training and separate validation tasks. The meta-agent does not see the validation tasks, their trajectories or their scores.
“We argue that interface design should itself be an agent capability.”
Saaket Agashe et al., “Evolving Agent Interfaces”
Qwen3.8-27B scores 21.3% on OSWorld V2
On OSWorld V2, Simular reports that Qwen3.8-27B with an evolved interface repository completed 21.3% of the benchmark’s 108 tasks under its binary success measure. The same model with fixed interfaces scored 5.56%, making the evolved setup’s result about 3.8 times higher. Simular compares the 21.3% score with 20.6% for Claude Opus 4.8 and 13.0% for GPT-5.5 in the company’s reported results.
Cost estimates in Simular’s table put the Qwen setup at $1.10 per task, compared with $72.40 for Claude Opus 4.8 and $25.50 for GPT-5.5. Those comparisons are not simply model-for-model tests: the listed Claude and GPT results use batched actions, while the Qwen row uses the evolved repository without a scratch virtual machine. The figures are Simular’s reported benchmark results and cost estimates, not an independent evaluation.
New interfaces reach beyond desktop chores
Simular says the evolved repositories added tools for system-wide actions, specific applications, work shared across applications and individual task procedures. Examples include running several GUI actions in one request, reading a file’s text and rendered pages without opening its original application, and checking whether a Lean proof compiles. In the OSWorld V2 run without a scratch virtual machine, the system added four interfaces and extended 127 existing ones; the run with a scratch VM added 12 interfaces and 253 extensions.
The company also tested Agent S4 on ScienceBoard, a 169-task benchmark spanning six scientific applications. With the same task agent and seed repository, the evolved version solved 99 tasks, compared with 93 for the seed. Simular says the meta-agent learned from 48 synthetic training tasks and did not see benchmark tasks.
The benchmark result has a narrow meaning
Agent S4 does not operate a desktop to complete a user’s task; its role is to develop and select the interfaces another agent uses. That distinction makes the system’s gains an argument for improving an agent’s available actions and observations alongside its underlying model. Simular has released the code, while the paper describing the method was submitted to arXiv on October 20, 2026.
The headline OSWorld figure also leaves most tasks unsolved: binary scoring counts a task as successful only when it passes, with no partial credit. Simular’s results show a marked jump over the same model with fixed interfaces, but they do not establish how reliably the method will transfer to other software, task sets or models. The code and paper give researchers a way to test that claim directly.