Draft:Agent S

• Comment: Besides the WP:NOLLM aspect, there's a lot of reliance on primary and non-independent sources (e.g. interviews, research papers from the same authors) which don't establish notability. Chaotic Enby (in solidarity · talk · contribs) 01:19, 3 November 2026 (UTC)

• Comment: In preparing this draft, I disclose that AI assistance was used as follows: The initial draft was written with Sai, an AI agent made by Simular (my employer, which also develops Agent S). I chose the sources from an earlier source review, and Sai wrote the wikitext from them. I read it through and asked Sai for changes, including plainer, less promotional wording and replacing self-published sources with conference listings where possible. I then checked the claims against the cited sources. The chat transcript is not publicly shareable. Samgyw (talk) 09:11, 2 November 2026 (UTC)

---

Agent S
Developer	Simular
Release	October 2024 (2024-10)
Written in	Python
Type	AI agent framework
License	Apache License 2.0
Repository	github.com/simular-ai/Agent-S

Agent S is an open-source framework for building computer-use agents. These are programs that operate a computer the way a person does, by reading the screen and using the mouse and keyboard. It is developed by researchers at the startup Simular. Four versions have been released since October 2024. Each is described in a research paper and tested on the OSWorld benchmark. The code is published on GitHub under the Apache License 2.0.

History

Agent S

Saaket Agashe, Jiuzhou Han, Shuyu Gan, Jiachen Yang, Ang Li and Xin Eric Wang posted the first Agent S paper to arXiv on 10 October 2024. It was accepted as a poster at the International Conference on Learning Representations (ICLR) in 2025. The Decoder covered the release that month and noted that the Python code was free to download from GitHub.

Agent S2

The Agent S2 paper followed on 1 April 2025. Wired covered it that month, focusing on how the agent switches between AI models depending on the task. The paper was accepted at the Conference on Language Modeling (COLM) in 2025. In July 2025, The Register reported that Agent S2 ranked fourth on the OSWorld leaderboard and sixth on AndroidWorld.

Agent S3

Agent S3 came out in October 2025, in a paper titled "Scaling Agents for Computer Use". The Chinese technology publication 机器之心 (Synced) reported that it was built directly on Agent S2, with a simpler framework and a built-in coding agent. The paper was accepted by the journal Transactions on Machine Learning Research (TMLR) in 2026.

Agent S4

The Agent S4 paper, "Evolving Agent Interfaces", was posted on 20 October 2026.

Architecture

The first version breaks a task into smaller steps before acting. To plan those steps, it searches for outside knowledge and looks up how it handled similar tasks before. Its authors call this experience-augmented hierarchical planning. It also uses an agent-computer interface (ACI), which the authors designed to help agents built on multimodal large language models reason about the screen and control it. An independent 2025 survey of GUI agents describes Agent S as turning user instructions into detailed subtasks, ordered so that each step comes after the steps it depends on.

Agent S2 splits the work between a general model and specialist models. The paper uses a technique called Mixture-of-Grounding to find elements on the screen precisely. A second technique, Proactive Hierarchical Planning, revises the plan at several time scales as the screen changes. Simular co-founder Ang Li told The Register that once the agent finds a way to do a task, it turns those actions into symbolic code that runs the same way each time. When that code stops working, the agent goes back to the language model.

A computer-use agent can succeed at a task on one run and fail on the next. Agent S3 runs several attempts in parallel and keeps the best one. Its selection method, Behavior Judge (BJudge), writes a short account of what each attempt did and compares those accounts. Agent S4 adds a meta-agent, a second agent that studies the task agent's attempts on practice tasks and builds a library of interfaces and notes on how to use them. Different versions of the library compete, and the ones that do best on tasks the meta-agent has not seen are developed further.

Benchmarks

On OSWorld, the first Agent S paper reported a success rate 9.37 percentage points above the baseline it was compared with. Synced reported OSWorld scores of 20.6% for Agent S and 48.8% for Agent S2. It gave 69.9% for Agent S3 at release, against a human score of about 72%. The current version of the Agent S3 paper reports 72.6% when several attempts are run, against a human score of 72.36%.

The Agent S2 paper also reported results on WindowsAgentArena and AndroidWorld, with relative improvements of 52.8% and 16.52% over the previous best methods. The Agent S4 paper reports a binary score of 21.3% on OSWorld V2 using the open-weight Qwen3.8-27B model.

Reception

The Decoder described Agent S as an attempt to take over routine office work such as data entry and scheduling. It also reported the developers' own finding that about 40 percent of failures came from weak task planning or from matching commands to the wrong screen elements. Wired wrote about Agent S2's use of more than one AI model. Synced covered Agent S3 at length. It named inconsistent results from one run to the next as a main obstacle for computer-use agents, and explained how running attempts in parallel reduces it.

Other researchers have used Agent S in their own work. The authors of OSWorld-Human, a benchmark that measures how efficiently computer-use agents work, studied Agent S2's runs on 39 OSWorld tasks to find where delays come from. As of 2 November 2026, the GitHub repository had about 12,850 stars and 1,500 forks.

External links

• Agent S on GitHub

Category:Free software programmed in Python Category:2024 software