What Agent S is
Agent S, from Simular, is an open-source (Apache-2.0) computer-use agent framework: it takes a natural-language task, looks at the screen and completes it by clicking, typing and scrolling in ordinary desktop and web apps, with no API integration or per-app scripting. On October 12, 2026 the repo had about 12,600 stars and recent commits (Oct 8), per the GitHub API. It runs on macOS, Windows and Linux and works with OpenAI, Anthropic and open-weight models.
Why it's in the news
Updates listed in the README:
- Oct 1, 2026: launch of the Simular API, a platform for building "always-on" computer agents. The company claims 90% token savings (vendor claim).
- Aug 28, 2026: Sai, Simular's hosted agent built on the repo's ideas, reportedly hit 73% on OSWorld 2.0, a 108-task benchmark of long tasks, ahead of GPT-5.6 Sol (62.57%, as reported by OpenAI) at lower cost. Result published by Simular itself.
- Jul 30, 2026: the Agent S3 paper ("Scaling Agents for Computer Use") was accepted to TMLR 2026.
- Agent S3 was the first to surpass human performance on OSWorld (72.60% vs ~72%), using Behavior Best-of-N. Alone, it scores 66% in the 100-step setting.
The numbers, in context
| Benchmark | Agent S3 alone | With Best-of-N |
|---|---|---|
| OSWorld | 66% | 72.6% |
| WindowsAgentArena | 50.2% | 56.6% |
| AndroidWorld | 68.1% | 71.6% |
Source: project README. Best-of-N runs several attempts and picks the best, which lifts the score but multiplies cost and isn't always feasible for tasks that change real state (sending an email has no "undo").
Getting started
pip install gui-agents
# on macOS, for OCR:
brew install tesseract
Then configure your model provider key. The project warns it's designed for a single monitor and runs Python code to control your computer, so use with care.
Where it fits in business
- Legacy systems without APIs: old ERPs, government portals, intranets.
- Repetitive back-office tasks: copying data between systems, building reports.
- End-to-end UI testing.
Risks and precautions
- Use a VM or isolated account, never your main desktop with email and banking open.
- Visual prompt injection: text on screen or a web page can instruct the agent. Limit reachable sites and apps.
- Human confirmation for irreversible actions (payments, sends, deletions).
- Credentials: don't leave saved passwords in the agent's session.
- Benchmarks aren't production: 72% success means roughly 1 in 4 tasks fails. Measure on your own workflow.
Sources
- Repository: https://github.com/simular-ai/Agent-S
- Agent S3 paper: https://arxiv.org/abs/2510.02250
- Simular, Sai on OSWorld 2.0: https://www.simular.ai/articles/sai-tops-osworld-2-0
- GitHub API (Oct 12, 2026)
Want to safely automate processes in systems without APIs? KGA IT offers a free consultation: kga-it.com.