{
  "id": 17052403,
  "title": "“Urgency is an inherently subjective call”: Anthropic tells TNS how it grades Claude Duet",
  "url": "https://urgent.news/2026/10/29/urgency-is-an-inherently-subjective-call-anthropic-tells-tns-how-it",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-29T11:00:00.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/claude-duet-agent-judgment/"
  },
  "original_language": "en",
  "account": "An AI assistant called Claude Duet, developed by Anthropic, monitors users' connected apps and works in the background to decide which tasks require attention without being explicitly instructed. After its debut on October 21, Anthropic has shared how it tests whether Duet prioritizes tasks correctly. The company has created internal graders to evaluate Duet's task prioritization, though urgency can be subjective and the graders are imperfect but helpful. Despite Duet's capabilities, it can still misjudge urgency or make mistakes when managing work across various apps. Duet can perform tasks like preparing work and flagging items requiring attention without requiring explicit instructions. However, it may not always determine which tasks should take priority. The company acknowledges the challenge of evaluating these decisions.\n\nOn October 23, two days after Duet's debut, Anthropic published a closer look at its safety, privacy, and security safeguards. Anthropic's Duet is separated from the part that checks if the work aligns with the user's intent, using a separate automated check built on the auto mode classifier from Claude Code and extended for Duet. This check ensures that instructions hidden in a webpage cannot directly influence the classifier's decision. Outside firms have tested Duet, but the company's public materials do not show how reliably the separate check catches actions outside the user's intent.\n\nDuet works with apps already connected to a user's Claude account, seeking approval before actions such as making payments or deleting items. Users cannot see everything happening behind the scenes, including Duet's cloud browser activity or taking over the browser. While users can review the task list and approve actions, they cannot inspect each browser step as it happens. Background work, like building a financial model or a deck, consumes resources from a subscriber's plan limits. Even routine checks across connected apps count toward the five-hour session and weekly limits, occurring every few hours to avoid exhausting a subscriber's allowance.",
  "summary": "Anthropic built Claude Duet, an AI assistant that monitors your connected apps and works on tasks in the background, to The post “Urgency is an inherently subjective call”: Anthropic tells TNS how it grades Claude Duet appeared first on The New Stack .",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}