{
  "schema_version": "1.0",
  "updated_at": "2026-10-23",
  "editorial_scope": "Evergreen explanatory topic page / evidence and linked news",
  "topic": "AI Agents",
  "canonical_url": "https://cortist.com/ai-agents/",
  "source_list": [
    {
      "title": "METR: Task-completion time horizons",
      "evidence_type": "Research / updated Sep 2026",
      "url": "https://metr.org/time-horizons/"
    },
    {
      "title": "METR: Frontier Risk Report",
      "evidence_type": "Research / May 2026",
      "url": "https://metr.org/blog/2026-05-19-frontier-risk-report/"
    },
    {
      "title": "METR: Limits of task-time measurements",
      "evidence_type": "Methodology / Jan 2026",
      "url": "https://metr.org/notes/2026-01-22-time-horizon-limitations/"
    },
    {
      "title": "Cortist calibration method audit",
      "evidence_type": "Internal review / Oct 2026",
      "url": "https://cortist.com/forecast/calibration/review/"
    }
  ],
  "linked_cortist_news": [
    {
      "slug": "openai-decisions-api-2026",
      "relationship": "PRODUCT LAUNCH",
      "warning": "The distinction between selecting an action and completing a whole workflow."
    },
    {
      "slug": "microsoft-project-perception-ai-agent-security",
      "relationship": "SAFETY",
      "warning": "Tool use and the importance of clear permission boundaries."
    },
    {
      "slug": "google-gemini-4-argon-developer-agents",
      "relationship": "PRODUCT LAUNCH",
      "warning": "Long-horizon agent capability claims need independent tests."
    }
  ],
  "primary_research_updates": [
    {
      "date": "September 8, 2026",
      "summary": "METR clarified that its public time-horizon page is no longer actively updated. Future evaluations require a comparable independent successor.",
      "url": "https://metr.org/time-horizons/",
      "source_type": "METR research"
    },
    {
      "date": "May 19, 2026",
      "summary": "METR described internal frontier-agent assessments and unusually long software-task results; private-system and benchmark conditions cannot be directly equated with public-agent outcomes.",
      "url": "https://metr.org/blog/2026-05-19-frontier-risk-report/",
      "source_type": "METR frontier report"
    }
  ],
  "forecast_candidate": "/forecast/ai-agents-eight-hour-tasks-2028/",
  "working_probability_is_independently_approved": false,
  "news_update_policy": "Manual editorial curation on publication. Do not automatically claim topical news when only external research updates exist."
}