Claude Duet Tested: Anthropic’s Work Agent, Helpful but Jumpy

Claude Duet is Anthropic’s new AI agent for work. An agent is software that carries out tasks on its own across several apps instead of only answering questions. According to The Verge, which published a hands-on test by Allison Johnson on November 13, 2026, Duet is “a decent helper, if a little jumpy at times”. Anthropic opened a research preview, meaning an early version that is still being tuned, in October 2026, according to its announcement.

What is Claude Duet and who can use it?

According to Anthropic, Duet is a standalone app on iOS and Android, and it also runs on claude.ai. During the preview it is included in the Pro, Max and Ultra plans and shares the usage limits of the rest of Claude. Access starts with subscribers in the United States and goes through a waitlist, according to Anthropic and The Verge.

Anthropic says it designed Duet around its Claude Opus 5.5 model, for “professionals and small business owners”. TechCrunch lists its rivals: Meta Muse, OpenAI Dots, Hark, Instinct, Grok Bot and Fo. The Verge describes Duet as even more focused on work than Muse, Dots or Instinct, and as more consumer-friendly than Claude Cowork, Anthropic’s earlier agent.

How does Duet work day to day?

Each morning Duet sends a short brief on the most important items across the connected apps, according to Anthropic. During the day it checks those accounts every few hours, because checking nonstop would use up plan limits much faster, and it messages the user only when something needs attention. In the app, The Verge adds, suggestions appear as cards that can be swiped left or right to hide or accept them, which its reviewer compares to a to-do list version of Tinder.

Anthropic says Duet has a virtual machine with its own browser, so it keeps working when the phone or laptop is off, and it sends a notification when a task is done or needs review. The user talks to it by typing or by voice in a single ongoing thread. Memory, skills and connected apps carry over from the user’s Claude account, according to Anthropic.

Duet can use more than 6,000 connectors and plugins from the Claude Marketplace, according to Anthropic and TechCrunch. Anthropic’s examples include:

• checking Intuit QuickBooks or Stripe for unpaid invoices and drafting reminders;
• preparing and running payroll in Gusto, with hours, bonuses and reimbursements filled in;
• drafting a job posting and, once the user edits and approves it, posting it to LinkedIn;
• sending agreements for review and signature through Docusign;
• drafting follow-up emails from the contact and deal history in HubSpot;
• turning a product launch into a social campaign in Canva;
• setting up and managing a Shopify store.

Anthropic cites an early tester named Amin, who owns the restaurant Shahs of Kabob in South Florida with about 110 employees. He used Duet for accounting, customer support and marketing, and worked through his invoices store by store. “I don’t need to open up 100 tabs,” he said, according to Anthropic.

What did The Verge’s hands-on test show?

The Verge’s reviewer gave Duet a work email account, a calendar, Google Docs and Slack. According to The Verge, Duet connected a point from a meeting summary email generated by Gemini to a direct message in Slack, and it worked out who the reviewer’s boss is from a shared Google Doc used for weekly one-to-one meetings.

In one case, Duet noticed that upcoming time off would leave only a couple of working days before a big draft was due, according to The Verge. It proposed a solution, offered to update the Google Doc with weekly notes and drafted a Slack message to the editor. The message copied the reviewer’s habit of writing in lowercase, and it carried a “sent with Claude” label.

In another case, the reviewer had told an interviewee that an article would likely go up on Tuesday. On Tuesday morning, before 9 AM, Duet pointed out that the story was still waiting for an edit. The Verge judges the worry misplaced: it was only 7 AM local time for the editor, and the message to the interviewee did not set a hard deadline.

The verdict is nuanced. The reviewer likes Duet as a backstop for their own organizational systems, because it finds discrepancies overlooked between emails, direct messages, meetings and shared documents. They do not want it as a replacement for those systems, and not in the creative process. The Verge also notes that AI in general is not good at understanding nuance or tone, and that Anthropic says it is still testing Duet’s judgment about when to message the user.

What limits and safeguards does Anthropic describe?

Before Duet sends a message, makes a payment, signs a document or deletes anything, it shows exactly what it is about to do and waits for the user’s approval, according to Anthropic. Some actions are off-limits even with approval, such as trading securities or moving money between accounts. The user can disconnect any app at any time.

Anthropic says Duet never directly sees passwords or sign-in tokens, following the design of Claude Managed Agents. Its actions are checked by the auto mode classifier already used for Claude Code. Against prompt injection, meaning hidden instructions planted in content the agent reads, Anthropic relies on its browser-use research. It says red teams and an external firm penetration tested Duet, and that it is covered by its public bug bounty program.

Anthropic also released a new classifier and a safety system prompt that check each message. By default they reject requests to share the strictest tier of information, including credit card numbers and government identification, according to Anthropic and TechCrunch. Health or personal finance data goes through only when the user explicitly instructs Duet to send it. For purchases, Duet can use Link, Stripe’s digital wallet, which pays with a virtual card so the real card number is not shared, or Shopify’s Shop Pay.

Anthropic is open about the limits of this early version. It says Duet might misread a thread or miscalibrate what is urgent, that personal errands such as booking tickets are weaker, and that it cannot yet show the user its browser. Duet lives in its own app, separate from the main Claude app. Its background checks count toward the plan’s usage limits, and larger tasks such as building a financial model or a deck need more model turns and background work, according to Anthropic.

What does this change for a company outside the United States?

Nothing can be deployed yet: according to Anthropic, the preview starts with Pro, Max and Ultra users in the United States, and the sources announce no date for other countries. For a company elsewhere, the case is mainly a preview of how proactive agents will work. This is our reading, not a claim from the sources.

Three questions are worth asking now. Which tools, such as email, calendar, chat and accounting, would you connect, and which data should never leave them? Which actions must always require a human approval? And how would you measure usefulness against the usage limits that background checks consume? Teams that want to frame these choices can start from our page on the agentic system.