🔍 Read the full analysis: Claude Duet’s Always-On AI Push Spotlights The Switchboard on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Anthropic launched Claude Duet on October 21 as a waitlisted research preview for US Pro and Max subscribers. The agent can work in the cloud across connected apps, but Anthropic says it requires approval for sensitive actions and blocks some financial activity; independent testing has not yet established how well its safeguards or features perform.
Anthropic launched Claude Duet on October 21 as a waitlisted research preview, introducing a cloud-based AI agent that can monitor connected apps, prepare work and message users when it needs their attention. The release adds Anthropic to a fast-moving competition over assistants that act between user prompts, while leaving performance and safety claims unverified by independent testing.
Duet is a separate app available through iOS, Android and claude.ai, according to Anthropic. It is built on Claude Opus 5.5 and is currently limited to US Pro and Max subscribers selected from a waitlist. Anthropic says it can send a morning brief, accept tasks by text or voice in an ongoing thread, and work on a virtual computer with its own browser while a user’s devices are offline.
The company describes possible tasks ranging from scheduling meetings and preparing dashboards to building sales forecasts or small web apps. Duet can use the memory, skills and connected apps associated with a user’s Claude account, and Anthropic says the Claude Marketplace offers more than 6,000 connectors and plugins. Its examples include checking invoices in QuickBooks or Stripe, payroll in Gusto, agreements through DocuSign and storefront tasks in Shopify. These are company descriptions; independent tests of the preview were not included in the source material.
Anthropic says Duet may read, draft and prepare work without waiting for each prompt, but sensitive actions require user approval. Before sending a message, paying or deleting something, the agent is supposed to show the action and its destination—including recipients—and wait. The company says trading securities, moving money between accounts and accessing another person’s account are blocked, even with approval. Anthropic also says purchases use a Stripe Link one-time card limited to a merchant and amount, rather than exposing the user’s regular card number.
Infographic 01 · The permission ladder
What Claude Duet does alone, what it asks first, what it never does.
Anthropic’s three tiers for an agent that keeps working while you are away. Health or financial details go out only with your explicit consent.
Does it alone
no approval
- Reads connected apps: mail, calendar, Slack, invoices
- Sends a morning brief, then flags what needs you
- Drafts replies, reminders, decks, dashboards
- Prepares payroll or contracts for your review
- Checks accounts every few hours, not nonstop
Asks you first
approval card shows action, destination, content, recipients
- Sending a potentially sensitive message
- Paying (one-time card via Link, limited to merchant and amount)
- Signing a document
- Deleting anything
- Never on “don’t ask me again”: paying or sending money, deleting an order, creating an account
Never, even if you approve
blocked
- Trading securities
- Moving money between accounts
- Accessing someone else’s account
- Sharing sign-in credentials, full card or bank numbers, government ID numbers
- Seeing your passwords, cookies or sign-in tokens (kept off the model’s machines)
“Approval fatigue is a serious issue, because too many requests teach people to tap through them.”Anthropic, Claude Duet safety post
These are Anthropic’s stated rules for the research preview. No independent test has been published yet.
Sources: Anthropic, Introducing Claude Duet (21 Oct 2026); Anthropic, Duet safety post (23 Oct 2026); Claude Duet FAQ.
Powered by Thorsten Meyer AI
Control Moves to the Agent Switchboard
Duet’s release is part of a broader shift from chatbots that answer on request to agents that can hold context, use connected services and act over time. That makes the software coordinating access to apps—not only the underlying language model—an increasingly important point of control. The service that manages a user’s connections, memory, permissions and task routing may shape what work gets done and which models perform it.
The supplied article argues that this “switchboard” could become a lasting competitive advantage as assistants route work to different models. It cites an October 6 post by Elon Musk saying SpaceX would use the best backend model for each task, including Claude Opus 5.5, and an xAI security FAQ stating that Grok Bot’s model choice cannot be restricted through the team model allowlist. Those examples point to a potential model-routing approach, but they do not establish how all competing products work or prove which company will control this layer.
For users and businesses, the immediate stakes are practical: what data an agent can access, what it can do without confirmation, and how to stop access. Anthropic says Duet checks connected accounts every few hours to manage usage, which may leave gaps between checks. The source also says data already learned by Duet remains until its memory is reset, even after an app is disconnected. Those details make permissions and data-retention policies central to any trial.
Products Worth Considering
Anker SOLIX C300 DC Portable Power Station with 60W Solar Panel, 288Wh, 300W
Samsung Odyssey G7 G75F 40-Inch WUHD Curved Gaming Monitor
LG UltraGear 32GX870A-B 32-inch 4K UHD OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
A Crowded Agent Launch Calendar
Duet follows several other agent announcements described in the source article. SpaceXAI opened a Grok Bot beta on August 11, Meta launched Muse on September 8, OpenAI introduced Dots at its DevDay on September 29, and Google released Gemini Agent on October 8, according to the article’s cited reporting. Anthropic announced Duet on October 21. The products are presented as sharing a general approach: an assistant works from a cloud environment and returns with results or requests for attention.
The launch information characterizes Duet as a research preview, not a fully established service. Anthropic’s safety post, dated October 23, gives more detail about approval prompts, prohibited actions and information handling than the initial product description. The source article reports that Anthropic says the agent does not see passwords, browser cookies or sign-in tokens, and that a classifier refuses to share credentials, full card or bank numbers, and government ID numbers. Health and financial details, Anthropic says, are shared only with explicit consent.
Anthropic’s own description also lists current shortcomings. Duet may misread a conversation or judge urgency incorrectly; personal errands such as booking tickets are weaker, and users cannot yet see its browser. Background checks count against the same usage limits as other Claude activity, while larger tasks can require more model turns. The source provides no independent comparison of Duet’s capabilities, availability or pricing against rival products.
“for professionals and small business owners with more on their plates than they can track”
— Anthropic, in its description of Claude Duet
Preview Limits and Safeguards
Independent testing is not yet available in the supplied reporting, so Duet’s real-world reliability, the effectiveness of its approval checks and the accuracy of its account monitoring remain unverified. Anthropic’s statements about what the agent can access and which actions it blocks are company claims, not findings from an outside audit described here.
It is also unclear how quickly the waitlist will expand, when Duet might leave research-preview status, or whether its features and access rules will change. Anthropic says app checks happen every few hours and that learned information can persist after an app is disconnected until memory is reset. The source does not specify the full retention period, how users can inspect all stored information, or how Duet’s safeguards compare with those of the other named agents.
Pilot Carefully, Await Broader Access
Anthropic has not supplied a public timetable for wider access or a general release. For now, the next practical milestone is whether users on the waitlist can test the research preview and report how it handles routine work, interruptions and approval requests. Any claims about performance should be treated as provisional until the product is examined beyond Anthropic’s examples.
The source article advises companies to run small, written pilots with limited, mostly read-only permissions and a defined exit plan. Users are advised to start with low-stakes tasks, connect only a few apps, and keep approval requirements enabled for anything that sends, pays or deletes. Those steps do not remove the risks described by Anthropic, but they limit exposure while the preview’s behavior and data controls are still being evaluated.
Key Questions
What is Claude Duet?
Claude Duet is Anthropic’s cloud-based AI agent, launched as a waitlisted research preview on October 21, 2026. It can work across connected apps and message users with briefs, task results or requests for attention, according to Anthropic.
Who can use Claude Duet?
The source says access is limited to US Pro and Max subscribers by waitlist. Anthropic has not provided a wider-release date in the supplied material.
Can Duet send messages or make payments on its own?
Anthropic says Duet can prepare work, but it must show sensitive actions and obtain user approval before sending, paying or deleting. The company says securities trading and transfers between accounts are blocked.
Has Duet’s safety been independently verified?
No independent test is described in the source material. The reported safeguards and restrictions are Anthropic’s claims, and their effectiveness in real-world use remains uncertain.
What should businesses do before trying an AI agent?
The source recommends a small, documented pilot with limited, mostly read-only permissions and an exit plan. Users should begin with low-stakes work and retain approval prompts for actions that send, pay or delete.
Source: ThorstenMeyerAI.com
