🔍 Read the full analysis: Claude Duet Joins The AI Agent Race, But Who Controls The Switchboard? on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Anthropic launched Claude Duet on Oct. 21 as a waitlisted research preview for U.S. Pro and Max subscribers. The company says the agent can work across connected apps from a cloud computer, but sensitive actions require approval and some financial actions are barred. Independent testing, broader availability and how Duet handles data already retained in memory remain open questions.
Duet is a separate app available through **iOS, Android and claude.ai**, according to Anthropic. The company says it runs on a virtual machine with its own browser, allowing it to continue tasks while a user’s phone or computer is offline. Users can give it requests by text or voice in an ongoing conversation, and Anthropic says it can send a morning summary and contact users during the day when attention is needed.
Anthropic says Duet carries over a user’s Claude **memory, skills and connected apps**, and can use more than 6,000 connectors and plugins in the Claude Marketplace. Its examples include finding unpaid invoices in QuickBooks or Stripe, working with payroll in Gusto, handling agreements through DocuSign, and managing a Shopify storefront. Anthropic also says purchases use Stripe’s Link wallet and a one-time card restricted to a single merchant and amount.
In a safety post dated **Oct. 23**, Anthropic said Duet presents the action, destination and content for review before sending sensitive messages, paying or deleting information. A separate classifier checks whether an action appears to match the user’s intent. Anthropic says trading securities, moving money between accounts and accessing another person’s account are prohibited, even with user approval. The company also says the agent does not see passwords, browser cookies or sign-in tokens.
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
Agents Put App Access in Focus
Duet’s launch adds another major technology company to a contest over agents that can act across a user’s software, rather than only answer prompts in a chat window. If these systems become dependable, they could take on routine business tasks such as tracking invoices, preparing reports and coordinating meetings. That could save time, but it also means giving an automated system access to work accounts and personal information.
The launch also points to a question beyond which model is most capable: **who controls the agent’s connections and task routing?** The source account says xAI’s Grok Bot can route work to models from other companies, including Claude Opus 5.5, while its security FAQ says customers cannot restrict Grok Bot’s model choice. If an agent selects a model behind the scenes, users and businesses may have less visibility into where a task goes and which provider processes it. The source material does not establish how Duet itself routes work to outside models.
For organizations, the practical issue is not only accuracy. Access permissions, approval habits, data retention and the ability to disconnect or leave a service all affect the risk of a pilot. Anthropic’s own warning about **approval fatigue** underscores that a review screen is not a complete safeguard if users become accustomed to accepting prompts without checking them.
Popular right now
TCL 85-Inch QM6K Series QLED Mini LED 4K HDR Smart Google TV
TCL 85 Designer Series QD-Mini LED TV (Model 85A400U) with Google Smart TV
Dyson Big Ball Multi Floor Canister Vacuum, Yellow
As an affiliate, we earn on qualifying purchases.
A Crowded Agent Launch Calendar
Anthropic’s Oct. 21 release followed several other launches described in the source account: SpaceXAI opened a Grok Bot beta on **Aug. 11**, Meta introduced Muse on **Sept. 8**, OpenAI announced Dots at DevDay on **Sept. 29**, and Google released Gemini Agent on **Oct. 8**, according to TechTarget as cited by the source. The products are presented as sharing a general approach: an assistant can work in the background on a cloud-based computer and return with a result or request for help.
Duet is based on **Claude Opus 5.5**, according to the launch material, and is not generally available. Its waitlist and plan restrictions make the release an early preview rather than a broad consumer rollout. The safety details currently available are Anthropic’s descriptions; the source says no independent test of Duet has been conducted.
The source also describes a shift toward agents that can choose among models. It attributes to Elon Musk an Oct. 6 post saying SpaceX would use the best backend model for a task, naming Claude Opus 5.5 and other services. It cites xAI’s FAQ language that Grok Bot’s team model allowlist does not govern the bot and that customers have no model picker. These statements concern Grok Bot, not a confirmed Duet capability.
““Approval fatigue is a serious issue, because too many requests teach people to tap through them.””
— Anthropic, in its Oct. 23 safety post
Testing and Data Retention Questions
There is **no independent evaluation** of Duet in the supplied source material, so its accuracy, security and performance in real work remain unverified. Anthropic says the preview may misread conversations or urgency, is weaker at personal errands such as booking tickets, and cannot yet show users its browser. The company also says it checks connected accounts every few hours, which may leave gaps between checks.
Anthropic says background activity counts against the usage limits shared with other Claude use, and that large tasks can require more model turns. It says disconnecting an app prevents new access, but information Duet has already learned remains in memory until the user resets that memory. The source does not specify how long other data may be retained, how the company will respond to errors during the preview, or when access may expand beyond U.S. Pro and Max subscribers.
It is also not clear how often Duet will ask users to approve actions in day-to-day use, or how well its separate intent check performs. Those details matter because the system’s safety relies partly on users noticing and reviewing prompts rather than accepting them automatically.
What Preview Users Should Watch
Anthropic has not provided a general release date in the supplied material. The next evidence will come from the research preview: whether Duet performs reliably across connected services, how users respond to its approval requests, and whether Anthropic changes its limits or access rules. Any claims about its practical accuracy should be treated as provisional until independent testing is available.
For businesses considering a trial, the source recommends **small, documented pilots**, read-mostly permissions and a clear exit plan. Users can start with low-stakes tasks, connect only a few apps and keep approval requirements on for actions that send, pay or delete. They should also understand how to reset Duet’s memory if they disconnect an account but do not want previously learned information to remain.
Any expansion of the waitlist, new supported services, changes to permitted actions or disclosures about data retention would help clarify whether Duet can move from an early preview into routine work. For now, Anthropic’s stated safeguards and limits describe the product, but they do not substitute for independent results.
Key Questions
Who can use Claude Duet now?
Anthropic says the research preview is **waitlist-only** and limited to subscribers in the United States on its Pro and Max plans.
What can Claude Duet do?
Anthropic says users can assign tasks by text or voice and connect work apps. Duet can prepare tasks on a cloud computer, send a morning brief and contact users when it needs attention. Its examples include invoice, payroll, contract and storefront work.
Can Duet send messages or spend money without permission?
Anthropic says Duet must show sensitive actions for approval before sending, paying or deleting. The company says trading securities, moving money between accounts and accessing someone else’s account are blocked.
Has Duet been independently tested?
The supplied source says **no independent test exists yet**. Anthropic has described the preview’s capabilities and limits, but those statements have not been independently verified in the source material.
What happens to information Duet has already learned if an app is disconnected?
Anthropic says disconnecting an app stops new access, but information already learned remains until the user resets Duet’s memory. The source does not provide further detail about retention periods.
Source: ThorstenMeyerAI.com
