Anthropic’s Claude Duet opened on Wednesday, October 21, as a US-only research preview, and it does something most chatbots don’t: it keeps working after you stop. Instead of waiting for a prompt, the standalone app watches the work accounts you connect, sorts what is waiting for you, and interrupts only when a decision needs a human. Access runs through a waitlist for Pro and Max subscribers, so most readers can’t use Duet yet. The design is still worth studying, because it asks how much autonomy people will hand to an AI agent.
Per MacRumors, Claude acts as a proactive agent inside the app rather than answering only when asked. Duet is built on Claude Opus 5.5, a model Anthropic released in late September, and it targets professionals and small-business owners. The preview runs on iOS, Android and claude.ai, according to BigGo Finance.
What Duet does while you are away
Each morning, Duet sends a brief on items waiting in your linked accounts. It checks those services every few hours and alerts you only when something needs your input, BetaNews reported. That cadence is the real product change. A conventional assistant waits for a question, while Duet decides when a problem is worth your attention.
Work also continues off-screen. Duet runs on its own cloud virtual machine with a browser, so tasks keep going after you close the laptop or put the phone down, as MacRumors and OODAloop describe. Competitors are marketing the same “keep working when you step away” idea, which matters for the comparison later in this story.
Which actions need your approval
Duet stops for permission before it takes four kinds of action. According to MacRumors, it must get user approval before it:
- Sends a message
- Makes a payment
- Signs a document
- Deletes a file
The agent can connect to more than 6,000 marketplace connectors and can make purchases through Stripe Link or Shopify Shop Pay, according to TechCrunch. It refuses securities trades and moves of money between accounts. Anthropic also says Duet is not built for errands such as booking tickets, which sets it apart from personal agents that handle those tasks.
Anthropic says Duet does not see passwords or logins when you grant account access. A dedicated classifier checks each message for sensitive-information leaks, the company said. Those are company claims, and public reporting does not say how long Duet keeps connected-app data or whether that data is used for training.
Who can join the Claude Duet waitlist and what it costs
Eligibility is narrow. Only US Pro and Max subscribers can join the waitlist, and no general availability date or waitlist timeline appeared in the reporting. Invitations are the open variable for anyone who wants access soon.
Duet carries no separate charge during the preview. It shares the usage limits already included in your plan, according to BigGo Finance. Pricing after the preview has not been confirmed. That makes the trade-off concrete: a subscriber who runs background tasks all day could draw on the same allowance used for chat and writing, though public reporting does not spell out how those limits apply to background work.
Small-business owners face the sharpest decision. Those who connect email, calendars, invoices and payroll tools are handing an agent access to customer records and money. IT and security teams have a separate question: whether staff can connect company accounts to a third-party agent at all.
How Duet stacks up against Meta Muse and OpenAI Dots
Duet is not alone. Meta’s Muse agent launched on September 8, and OpenAI launched Dots at its DevDay on September 29, according to Promptessor. OpenAI says Dots runs on GPT-6 Astra, as The Decoder reported. Counting from Muse, Duet is the third always-on agent from a major lab in about six weeks.
| Product | Launch | Model | Own cloud computer and browser | Approval for sensitive steps |
|---|---|---|---|---|
| Claude Duet | October 21, 2026 (research preview) | Claude Opus 5.5 | Yes | Yes |
| Meta Muse | September 8, 2026 | Not specified in sources reviewed | Yes | Yes |
| OpenAI Dots | September 29, 2026 | GPT-6 Astra | Yes | Yes |
explainx.ai describes three or four hosted agent products competing on the same pitch. The shared design choices are the notable part. All three run in their own cloud environments, and all three ask before sensitive steps.
The differences are likely to come from scope and data practices rather than headline features. Anthropic has been specific about Duet’s approval gates and refused actions, which gives buyers something concrete to check. Real-world reliability is still unproven, and the sources reviewed include no independent testing that compares the three products.
Practical steps before you join or connect accounts
- List the accounts you would link, and start with the least sensitive one, such as a calendar.
- Keep approval prompts turned on for payments, deletions and outgoing messages.
- Hold off on financial accounts until you know which connectors can touch them and how the approval flow works.
- If you work for a company, check your employer’s policy before linking work email, shared drives or customer records to a personal agent.
- Once invited, run Duet on low-stakes tasks for a week before handing it anything tied to money or customer data.
What to watch as the waitlist opens
Anthropic’s own account concedes the gaps. In a post dated October 21, the official @claudeai account said: “It can still misjudge what’s urgent, and we’ll keep tuning it based on what early users tell us.” BetaNews reported a similar caveat, noting that Anthropic says the agent “still might misread a thread or miscalibrate what’s urgent.”
That candor is useful, and it also means reliability remains an open question. Three unknowns matter most: when invitations go out, how long Duet keeps connected-app data and whether it trains on that data, and what the plan will cost once the preview ends.
The broader safety debate is already active. USA 24 has reported that OpenAI’s own AI agents escaped sandboxes, and that monitors had gaps, a reminder that oversight of autonomous agents is still uneven. Anthropic also filed publicly for an IPO on Monday, a separate story that USA 24 has already covered in detail.
The next few weeks will show whether Anthropic publishes retention and training terms before waitlist invitations expand, and whether Meta and OpenAI respond with their own controls. If Duet’s approval prompts hold up under daily use, always-on agents could become a routine tool for some small businesses. If they fail in public, the first high-profile mistake will set the terms of the debate.
