{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-01-026",
    "source_story_id": "tmp-story-osworld-science-agents",
    "edition_id": "mp-2026-10-01-morning-0084",
    "edition_url": "https://themachinepress.com/edition/2026-10-01",
    "position": 15,
    "story_type": "dispatch",
    "section": "benchmarks-evals",
    "editorial_classification": "editorial",
    "headline": "Scientific Software Exposed the Limits of Computer-Use Agents",
    "slug": "scientific-software-exposed-the-limits-of-computer-use-agents",
    "dek": "OSWorld-Science judged generated molecules, masks, plots and numbers—not just screenshots or final clicks.",
    "summary": "OSWorld-Science judged generated molecules, masks, plots and numbers—not just screenshots or final clicks.",
    "body_text": "OSWorld-Science provides 146 expert-designed tasks across molecular drawing, retrosynthesis, pathology imaging, statistics and physical simulation. Execution-based evaluators inspect application state and artifacts such as molecular structures, segmentation masks, plots and numerical results, awarding partial credit when work is incomplete. Tests of twelve vision-language models with a strong harness still found substantial difficulty. The benchmark connects agent action to scientific outputs, but it does not establish that its automated checks replace domain-expert review.",
    "why_it_matters": "OSWorld-Science judged generated molecules, masks, plots and numbers—not just screenshots or final clicks.",
    "limitations": [
      "The benchmark connects agent action to scientific outputs, but it does not establish that its automated checks replace domain-expert review."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-01-026/scientific-software-exposed-the-limits-of-computer-use-agents",
    "json_url": "https://themachinepress.com/story/mp-2026-10-01-026.json",
    "first_published_at": "2026-10-01T09:00:00.000-04:00",
    "modified_at": "2026-10-01T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-01-026-001",
        "text": "OSWorld-Science judged generated molecules, masks, plots and numbers—not just screenshots or final clicks.",
        "source_ids": [
          "source-2026-10-01-015"
        ],
        "qualification": "The benchmark connects agent action to scientific outputs, but it does not establish that its automated checks replace domain-expert review."
      }
    ],
    "source_ids": [
      "source-2026-10-01-015"
    ],
    "tags": [
      "computer use",
      "scientific software",
      "artifact evaluation"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-01-015",
      "title": "arXiv preprint 2609.39903",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.39903",
      "canonical_url": "https://arxiv.org/abs/2609.39903",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-30T11:03:05.000-04:00",
      "accessed_at": "2026-10-01T08:18:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-01-026-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Scientific Software Exposed the Limits of Computer-Use Agents",
    "publisher": "The Machine Press",
    "published_at": "2026-10-01T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-01-026/scientific-software-exposed-the-limits-of-computer-use-agents"
  }
}