Agent S — Open Computer-Use Agent (S1 to S4, Real GUI Control on macOS, Windows and Linux)

Agent S is an open-source computer use agent framework from Simular. Give it a task in plain English. It looks at the screen and completes the task by clicking, typing, and scrolling in ordinary desktop and web apps. It runs on macOS, Windows, and Linux.

🥳 Updates

• 2026/10/20: Released Agent S4 and its paper "Evolving Agent Interfaces": an autonomous computer-use scientist that evolves the interfaces its agents work through from task experience. With Qwen3.8-27B, Agent S4 sets a new open-source state of the art on OSWorld V2 at 21.3%, 3.8 times the same model with fixed interfaces and above GPT-5.5 (13.0%) and Claude Opus 4.8 (20.6%) at a fraction of their cost. (OSWorld V2 is scored as a strict binary pass/fail with no partial credit, so 21.3% reflects benchmark difficulty, not low reliability — see the Agent S4 README for details.)
• 2026/10/01: The Simular API is live. Build your always-on computer agents in a minute. Simular Platform is an autonomous computer API for building always-on agents that fit your needs. World-leading agent infrastructure, 90% token saving. Get an API key →

• 2026/08/28: Sai tops OSWorld 2.0, beating GPT and Opus with lower costs (73%) — our production computer use agent is now state of the art on the hardest desktop benchmark!
• 2026/07/30: The Agent S3 paper "Scaling Agents for Computer Use" is accepted to TMLR 2026!
• 2025/12/15: Agent S3 is the first to surpass human-level performance on OSWorld with an impressive score of 72.60%!
• 2025/10/02: Released Agent S3 and its technical paper, setting a new SOTA of 69.9% on OSWorld (approaching 72% human performance), with strong generalizability on WindowsAgentArena and AndroidWorld! It is also simpler, faster, and more flexible.
• 2025/08/01: Agent S2.5 is released (gui-agents v0.2.5): simpler, better, and faster! New SOTA on OSWorld-Verified!
• 2025/07/07: The Agent S2 paper is accepted to COLM 2025! See you in Montreal!
• 2025/04/27: The Agent S paper won the Best Paper Award 🏆 at ICLR 2025 Agentic AI for Science Workshop!
• 2025/04/01: Released the Agent S2 paper with new SOTA results on OSWorld, WindowsAgentArena, and AndroidWorld!
• 2025/03/12: Released Agent S2 along with v0.2.0 of gui-agents, the new state-of-the-art for computer use agents (CUA), outperforming OpenAI's CUA/Operator and Anthropic's Claude 3.7 Sonnet Computer-Use!
• 2025/01/22: The Agent S paper is accepted to ICLR 2025!
• 2025/01/21: Released v0.1.2 of gui-agents library, with support for Linux and Windows!
• 2024/12/05: Released v0.1.0 of gui-agents library, allowing you to use Agent-S for Mac, OSWorld, and WindowsAgentArena with ease!
• 2024/10/10: Released the Agent S paper and codebase!

🚀 Quick Start

• No setup: run Sai, our production agent, on a cloud computer through the Simular API. See the API docs and the workflow template gallery.
• On your own computer: follow the Agent S3 README.

📚 Agent S Versions

Version	What it is	Docs	Paper
Agent S4	Evolves the tool interfaces a computer use agent works through, from its own task experience. Open-source state of the art on OSWorld V2 (21.3% with Qwen3.8-27B).	README	arXiv 2026
Agent S3	The agent in gui-agents today. First to surpass human performance on OSWorld (72.60%).	README	TMLR 2026
Agent S2.5	A simpler, faster Agent S2.	README	—
Agent S2	A compositional generalist-specialist framework.	README	COLM 2025
Agent S	The first version: uses computers like a human, through an agent-computer interface.	README	ICLR 2025

💬 Citations

If you find this codebase useful, please cite:

@misc{Agent-S4,
      title={Evolving Agent Interfaces}, 
      author={Saaket Agashe and Yuzhe Yang and Gonzalo Gonzalez-Pumariega and Bhuvan Chandra Koduru and Chih-Lun Lee and Jiachen Yang and Ang Li and Xin Eric Wang},
      year={2026},
      eprint={2610.28601},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.28601}, 
}

@misc{Agent-S3,
      title={The Unreasonable Effectiveness of Scaling Agents for Computer Use}, 
      author={Gonzalo Gonzalez-Pumariega and Vincent Tu and Chih-Lun Lee and Jiachen Yang and Ang Li and Xin Eric Wang},
      year={2025},
      eprint={2510.02250},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2510.02250}, 
}

@misc{Agent-S2,
      title={Agent S2: A Compositional Generalist-Specialist Framework for Computer Use Agents}, 
      author={Saaket Agashe and Kyle Wong and Vincent Tu and Jiachen Yang and Ang Li and Xin Eric Wang},
      year={2025},
      eprint={2504.00906},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2504.00906}, 
}

@inproceedings{Agent-S,
    title={{Agent S: An Open Agentic Framework that Uses Computers Like a Human}},
    author={Saaket Agashe and Jiuzhou Han and Shuyu Gan and Jiachen Yang and Ang Li and Xin Eric Wang},
    booktitle={International Conference on Learning Representations (ICLR)},
    year={2025},
    url={https://arxiv.org/abs/2410.08164}
}

Star History

---

TalentDelta provenance record (added at packaging)

• Packaged for: TalentDelta catalogue (draft; gated by review)
• Upstream: simular-ai/Agent-S
• Pinned commit: 0fff0e634869dab72027435b4586e62300185d41 (2026-10-21)
• Licence: Apache-2.0 — Apache License 2.0 at the repository root, shipped byte-identical to the pin. Permissive: use, modification, distribution and commercial use allowed with attribution, licence text and change notices. No NOTICE file exists upstream; no per-file headers are added and no file is modified. Attribution to Simular is recorded in the pack README's provenance section and in this upload record.
• Upstream author credit: Simular (simular-ai)
• Runtime: Python 3.11+ with platform extras (pip install -r requirements.txt; pyobjc on macOS, pywinauto/pywin32 on Windows); GUI control libraries pyautogui, pytesseract and PaddleOCR; model backends through OpenAI, Anthropic, Google GenAI or Together keys; optional OSWorld / WindowsAgentArena VMs for evaluation
• Changes made at packaging: this section was appended to README.md. No upstream path is excluded. Apart from this provenance section appended to README.md, no file was added, removed, renamed or edited; every upstream file is byte-identical to the pinned commit.
• Safety read of the packaged archive (packaging time, 2026-10-24): Static safety read of the shipped archive (356 files): 75 pattern hits, none a live credential. 40 are api-key patterns, all of them environment-variable reads or client-construction arguments with no literal value; 31 are home-directory placeholders inside the VM provisioning scripts; 2 are addresses - the maintainer's in setup.py metadata and an example address in a test; 2 are absolute user paths from upstream, one a default benchmark example directory. No .env file ships.

Packaging notes (TalentDelta)

• A working agent, not a prompt set: each generation ships an agent class, grounding model wiring, action loop, memory and worker code under gui_agents/ and agent_s4/.
• Computer use is powerful by construction: the agent clicks and types in real applications with the operator's own session. The shipped README and agent READMEs describe the sandbox/VM route for evaluation; the pack adds no elevation and no credential handling of its own.
• No API keys or model weights ship. Keys are read from the operator's environment; requirements.txt names the vendor SDKs only.
• Upstream benchmark claims (OSWorld, WindowsAgentArena, ScienceBoard) are quoted from its README and papers; this pack re-runs none of them and asserts no score.
• Screenshots in images/ are upstream product and benchmark figures, not captures of any user's desktop.