🔍 Read the full analysis: How The Switchboard Shapes The Always-On AI Agent Race on ThorstenMeyerAI.com
Get the little things that make your day delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Anthropic launched Claude Duet as a waitlisted research preview on October 21, joining a growing field of AI agents that work on cloud computers and return with results or requests for approval. Anthropic says sensitive actions require user approval and some financial actions are blocked, but independent testing has not been reported. The wider competition is also about who controls connected apps, stored context and the choice of model behind each task.
Anthropic launched Claude Duet on October 21 as a waitlisted research preview, adding a cloud-based AI agent to a fast-growing field that includes products from OpenAI, Meta, SpaceXAI and Google. The app is designed to work across connected services while a user’s devices are offline, bringing attention to who controls the agent’s app access, user context and model routing.
Duet is a separate app available through iOS, Android and claude.ai. Anthropic says it runs on a virtual machine with its own browser and can continue working while a user’s phone or computer is off. The company describes a daily brief, tasks sent by text or voice, and notifications when a decision or other attention is needed. The preview uses Claude Opus 5.5 and is limited to US Pro and Max subscribers selected through a waitlist.
Anthropic says Duet can use memory, skills and connected apps from a user’s Claude account, including services for accounting, payroll, contracts and online stores. The company cites more than 6,000 connectors and plugins in the Claude Marketplace. Its examples include checking unpaid invoices in QuickBooks or Stripe and working with payroll in Gusto. Those capabilities and descriptions come from Anthropic; the supplied source material does not include independent testing of the preview.
In a safety post published October 23, Anthropic said Duet must show users a proposed action and wait for approval before it sends a potentially sensitive message, makes a payment or deletes something. The company says trading securities, transferring money between accounts and accessing another person’s account are blocked. It also says Duet does not see passwords, browser cookies or sign-in tokens. These are company statements about safeguards, not independently verified findings.
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 of Agents and Their Connections
The products in this race are being positioned not just as chatbots, but as agents that can act across services and return with completed work. That makes the connection layer important: an agent’s usefulness depends on what accounts it can access, what information it retains and which actions it is allowed to take. For businesses, the same access creates questions about oversight, data handling and the consequences of mistakes.
The source report also points to a contest over the switchboard: the layer that routes a task to a model and connects it to apps and user memory. It cites an October 6 post by Elon Musk saying SpaceX would use the best backend model for a task, including models from other companies. It also cites xAI’s security FAQ, which says the team model allowlist does not govern Grok Bot and that customers do not have a model picker. If that description holds in practice, users may interact with one assistant while a different company’s model handles particular work.
That model-routing shift could make the service controlling the interface and connections more important to a customer’s day-to-day experience than any single underlying model. It is an interpretation of the product direction described in the source, not a confirmed measure of market power or a settled outcome. For customers, it makes data portability, permission controls and visibility into routing relevant questions when selecting an agent.
Products Worth Considering
Philips Hue Large Smart Light Tube, Black – White and Color Ambiance LED Color-Changing Light – 1 Pack
Blueair Blue Pure 211i Max Smart Air Purifier for Large Rooms
Samsung 42-Inch S90H OLED 4K Smart TV Bundle
As an affiliate, we earn on qualifying purchases.
A Crowded Agent Launch Calendar
The source material dates the recent launches across several months: SpaceXAI opened the Grok Bot beta on August 11, Meta launched Muse on September 8, OpenAI introduced Dots at its September 29 DevDay, and Google released Gemini Agent on October 8, according to TechTarget as cited in the report. Anthropic announced Duet on October 21. The products are presented as sharing a broad approach: agents operate beyond a single chat exchange, using cloud computing and connected services to perform tasks over time.
Anthropic’s own account also describes limits in the preview. Duet may misread a message thread or judge urgency incorrectly; personal errands such as booking tickets are weaker, and users cannot yet watch its browser. Connected accounts are checked every few hours, the company says, partly to manage plan usage. Background activity draws on the same usage limits as other Claude work, and large assignments can require more model turns.
Those details matter because an agent that runs in the background is not simply a faster chat interface. It needs ongoing access and can use account limits even when a user is not actively prompting it. Anthropic says disconnecting an app stops future access, but information Duet has already learned remains until its memory is reset. That distinction is relevant to users evaluating retention and account disconnection.
““Approval fatigue is a serious issue, because too many requests teach people to tap through them.””
— Anthropic, in its October 23 safety post
Preview Safeguards Still Need Testing
Independent test results for Duet’s accuracy, permissions and safety controls are not included in the source material. Anthropic has described approval steps and prohibited actions, but it remains unclear how reliably the system identifies sensitive actions or interprets a user’s intent in real use. The company itself warns that users may face approval fatigue.
It is also not clear how Duet’s capabilities, availability or limits will change after the waitlisted preview, or whether access will expand beyond eligible US Pro and Max subscribers. The source report gives launch-era descriptions and pricing context but does not establish how those terms may change. For the wider market, claims about switchboard control remain a developing interpretation: the material does not provide independent evidence about how often agents route tasks to outside models or what customers can see when they do.
Anthropic says Duet checks connected accounts every few hours, but the source does not quantify how often that schedule could cause missed updates. It also says learned information persists after an app is disconnected until memory is reset; the exact retention behavior and how users can review what was retained are not detailed in the material.
Watch for Preview Changes and Tests
The immediate next step is the expansion and evaluation of Duet’s waitlisted research preview. Users and businesses will need to test how it performs on routine tasks, how often it requests approval, and whether its safeguards are clear in practice. Anthropic’s published feature descriptions do not establish how the agent performs under independent review.
Companies considering a trial can limit the initial scope, document which accounts are connected and use read-mostly permissions where possible. The source report recommends small written pilots with an exit plan. Individual users can begin with low-stakes work, connect only a few services and retain approval requirements for actions that send, pay or delete. These are precautions, not a substitute for independently verifying the product’s controls.
Further developments to watch include changes to Duet’s availability and capabilities, the frequency of its background checks, and whether services disclose which models handle particular tasks. Those details will help clarify whether control of the interface, app connections and routing layer becomes a practical advantage for one provider or remains a feature of a crowded market.
Key Questions
What is Claude Duet?
Claude Duet is Anthropic’s separate, cloud-based agent app, announced as a waitlisted research preview on October 21, 2026. Anthropic says users can send it tasks by text or voice and receive briefs or messages when attention is needed.
Who can access the preview?
According to Anthropic’s launch details in the source report, Duet is limited to US Pro and Max subscribers selected through a waitlist. Wider availability has not been confirmed in the supplied material.
Can Duet take actions without approval?
Anthropic says Duet can read, draft and prepare work, while potentially sensitive messages, payments and deletions require approval. The company says securities trading and transfers between accounts are blocked. These safeguards have not been independently tested in the source material.
What does the “switchboard” mean in this report?
It refers to the service layer that connects an agent to a user’s apps and context and may choose which model handles a task. The source report argues that control of this layer could matter as much as the underlying model, but that is an interpretation rather than a confirmed market outcome.
Source: ThorstenMeyerAI.com
