October 21, 2026 · Research
Agent S4: evolving the interfaces computer-use agents work through
21.3% on OSWorld V2 with an open-weight Qwen model at about USD 1.10 per task — on par with Claude Opus 4.8 at roughly 1.5% of the cost.
OUR JOURNEY
A timeline of discovery.
- 2026
October 16, 2026 · Research
Put the Frontier Model on a Leash
LEASH lets a small open-weight agent call a frontier model only when uncertain: 73.66% on OSWorld, matching always-on Claude Opus 5 at lower cost.
October 8, 2026 · Research
The model says no in chat. The agent does it anyway
Refusals drop sharply when models move from chat to controlling a desktop. Sai declined 69% of OS-Harm misuse tasks on the first turn.
October 1, 2026 · Platform
Simular API
Our research, now an API: give any AI agent a real computer to see, click and type. Run once, replay as code — with 90% token savings.
October 1, 2026 · Research
Neither code nor LLMs: Real work needs self-healing agents
Neuro-symbolic agents learn a task once, then replay it as self-healing code — up to ~200× lower cost than running base models.
September 21, 2026 · Perspective
What Is a Robosecretary?
A computer-use agent that works inside your existing software and replays solved tasks as code across a fleet of computers.
September 16, 2026 · Product
Your first robosecretary has arrived
Sai is generally available: delegate desktop tasks on cloud or personal machines and get notified when the work is done.
August 28, 2026 · Research
Sai tops OSWorld 2.0, using least tokens
Beats GPT-5.6 and Claude Opus on a 108-task benchmark that takes skilled humans over an hour.
August 21, 2026 · Open Source
Sai-Fi: voice interface for our robosecretary Sai
Prototyped on Meta Ray-Ban glasses. Talk to the computer-use agent hands-free, hand off desktop tasks, and interrupt them mid-run.
August 12, 2026 · Research
Co~work~play: extending computer-use to real-time gameplay
New harnesses let agents measure the screen and act in real time; with Scenes, Sai scored ~67% where frontier models stayed near 20%.
July 30, 2026 · Research
Agent S3 paper accepted to TMLR
Our “Scaling agents for computer use” paper accepted by Transactions on Machine Learning Research
June 17, 2026 · Perspective
The Power Law of Practice Is Absent From Agents
Agents lack the practice-makes-perfect learning curve. A neuro-symbolic compilation approach is the future.
May 19, 2026 · Partnership
Windows 365 for Agents enters enterprise workflows
Extends Sai into enterprise agent workflows.
May 15, 2026 · Open Source
Playwright for the entire desktop
A scripting language for automating browsers, native apps, and OS-level workflows
April 23, 2026 · Product
Sai × Claude Code
Lets the agent drive coding tools autonomously — no babysitting required.
February 2026 · Perspective
The Unity of Opposites in Building Computer-Using Agents
Human vs machines, natural language vs programming language, entropy vs order, and flexibility vs reliability
February 2026 · Perspective
The Case for GUI Agents
GUI agents are the last mile of computer intelligence
January 26, 2026 · Perspective
The Year Desktop Agents Stop Being a Toy
Agentic reliability finally gets the right eval
- 2025
December 15, 2025 · Research
Agent S3 surpassed human performance on OSWorld
First achievement in the industry. By our open-sourced agentic framework.
December 2, 2025 · Company
Simular raises a $21.5M Series A funding round
From investors including Felicis, NVIDIA’s NVentures, Basis Set, and more.
November 18, 2025 · Partnership
Invited by Microsoft to pilot Windows 365 for Agents
Microsoft’s new infrastructure to help developers create computer and browser-based AI systems
October 2, 2025 · Research
Agent S3 approaching human performance
Introduced a wide-scaling framework for computer-use agents through Behavior Best-of-N (bBoN)
July 8, 2025 · Product
Simular Pro release
The first production-grade computer use agent in the industry
July 7, 2025 · Research
The Agent S2 paper is accepted to COLM 2025
May 2025 · Research
Agent S wins Best Paper Award at the ICLR 2025
Awarded by the Agentic AI for Science Workshop
April 1, 2025 · Research
Agent S2 set SOTA on OSWorld, WindowsAgentArena and AndroidWorld
Ahead of OpenAI and Anthropic
March 27, 2025 · Partnership
Selected for the Google Cloud AI Accelerator
Supercharging Simular’s ability to scale the brain for autonomous computers on cutting-edge cloud infrastructure
January 2025 · Research
Agent S paper accepted to ICLR 2025
Establishes the computer-use agent framework academically.
- 2024
December 5, 2024 · Research
gui-agents v0.1.0 released
Agent S opens up as a pip library for the first time — anyone can run it on their own machine.
October 9, 2024 · Research
Agent S paper and framework released
The first open agentic framework that uses a computer like a human: experience-augmented hierarchical planning + retrieval-based memory + an Agent-Computer Interface.
- 2023
August 7, 2023 · Company
Simular founded
Building accessible general intelligence to liberate human labor.
OUR LATEST THINKING