---
title: 'OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments'
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is CC BY 4.0 (verified on arXiv abs page 2026-11-11; icon by-4.0.png, http://creativecommons.org/licenses/by/4.0/); note is our own summary'
source: https://iinventedlight.com/library/notes/2404-07972-osworld.md
source_license: 'CC BY 4.0'
attribution: 'Summary by iinventedlight (AI-assisted). Source: Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, et al. (17 authors), OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments (CC BY 4.0).'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-11
status: published
approved_by: "Cole Withington, 2026-11-11 20:58 PT, Grok Bot chat widget t1717s7407: \"Publish all 100 including the LLMs Encode Clinical Knowledge note\""
checked_on: 2026-11-11
as_of: 2026-11-11
confidence: medium
untested: 'Not reproduced. Abs page and HTML key sections read 2026-11-11. Named code URLs checked; code not run.'
category: arxiv
topic_area: 'ai-research: agents'
tags: arxiv, agents, computer-use, evals
arxiv_id: 2404.07972
alias: 2404.07972-osworld
sources:
  - id: S1
    title: 'OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments'
    publisher: arXiv
    authors: 'Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, et al. (17 authors)'
    url: https://arxiv.org/abs/2404.07972
    version_read: 'v2 (30 May 2024)'
    accessed: 2026-11-11
    license: 'CC BY 4.0'
---
# OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments

## Summary
OSWorld is a real computer environment for multimodal agents, with task setup, execution-based evaluation, and interactive learning across Ubuntu, Windows, and macOS. The authors build a benchmark of 369 computer tasks involving real web and desktop apps and multi-app workflows. They report human success over 72.36% and best-model success of only 12.24%. [S1]

## Claims
- id: C1
  claim: 'OSWorld is a real computer environment for multimodal agents across Ubuntu, Windows, and macOS.'
  source: S1
  evidence: "we introduce OSWorld, the first-of-its-kind scalable, real computer environment for multimodal agents"
- id: C2
  claim: 'The benchmark has 369 computer tasks involving real web and desktop apps.'
  source: S1
  evidence: "a benchmark of 369 computer tasks involving real web and desktop apps"
- id: C3
  claim: 'They report human success over 72.36% and best-model success of 12.24%.'
  source: S1
  evidence: "humans can accomplish over 72.36% of the tasks, the best model achieves only 12.24% success"

## Verification
- Code available: The paper points to https://os-world.github.io, which returned HTTP 500 from this fetcher on 2026-11-11, so availability of that project page was not verified here. Not run.
- Reproduced: not reproduced (2026-11-11). Code not run.
- Related work: Live notes ao-evals-agentbench-osworld and ao-evals-multimodal-agents. This is the paper-verification note for OSWorld itself.
- Limits: The 12.24% best-model figure is from the v2 evaluation (revised 30 May 2024) and does not reflect later computer-use models. The official project URL named in the paper did not return HTTP 200 to this fetcher.
- License: CC BY 4.0 on the abs page (2026-11-11; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 11 Apr 2024 (v1); last revised 30 May 2024 (v2). Sources: [S1] https://arxiv.org/abs/2404.07972, v2 (30 May 2024). checked_on: 2026-11-11.

## When to apply
Use when evaluating computer-use or desktop agents on real applications, not on a toy mock UI.

## Facts
- id: F1
  claim: 'OSWorld is a real computer environment for multimodal agents, covering Ubuntu, Windows, and macOS.'
  confidence: medium
  date: 2026-11-11
  sources: [S1]
  license: CC BY 4.0
- id: F2
  claim: 'The authors build a benchmark of 369 computer tasks involving real web and desktop apps and multi-app workflows.'
  confidence: medium
  date: 2026-11-11
  sources: [S1]
  license: CC BY 4.0
- id: F3
  claim: 'They report human success over 72.36% and best-model success of 12.24% in their evaluation.'
  confidence: medium
  date: 2026-11-11
  sources: [S1]
  license: CC BY 4.0

## Sources
- [S1] arXiv 2404.07972, https://arxiv.org/abs/2404.07972. License: CC BY 4.0. v2 (30 May 2024). Accessed 2026-11-11.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

---

Agents: [/agents.md](/agents.md) | [/.well-known/agent-library.json](/.well-known/agent-library.json).