Claude Duet: An agent that already knows in the morning what is on fire – and asks before sending

Imagine your inbox getting in touch with you in the morning, instead of the other way round. A short overview of the key points, a note about the appointment that does not fit your flight, and the question of whether a reply to a customer may go out as written. This is exactly how Anthropic describes its new app "Claude Duet", which was presented on October 21, 2026. According to the vendor, it is aimed at self-employed people and small businesses and is meant to complete tasks even when you are not at your computer.

For managing directors of mid-sized companies, the announcement is interesting for two reasons. First, it shows where assistants are heading: away from a chat window into which you type something, toward an agent that notices things on its own. Second, the vendor builds an approval step in front of critical actions. That is precisely the question that decides, in practice, whether an agent helps you or causes you trouble. One important point up front: the product is currently a research preview with a waitlist and launches with subscribers in the US. This article therefore treats it as something that is coming toward businesses, not as something you can switch on tomorrow.

What Anthropic presented on October 21, 2026

The announcement is on the Claude blog (claude.com/blog, "Introducing Claude Duet", October 21, 2026, accessed on October 22, 2026). According to it, Claude Duet is a standalone app for iOS and Android, also reachable via claude.ai, separate from the existing Claude app. The vendor describes it as Claude in the role of a proactive agent: it points out things that need your attention, completes tasks you give it, keeps working when you are away, and asks before it acts on your behalf.

The page names the following specific features:

• Morning overview: a daily summary of the key points from the connected apps.
• Watching: The agent checks the connected apps over the course of the day, every few hours according to the page, and only gets in touch when something needs attention.
• Tasks: According to the page, the range extends from scheduling appointments to dashboards, financial or sales forecasts, prototypes and web apps.
• Working without your device: Duet runs on a virtual machine with its own browser; your laptop or phone does not need to be switched on.
• Shared context: Memory, skills and connected apps are carried over from the Claude account.
• Operation: a single continuous conversation thread, by typing or by voice.
• Integrations: access to more than 6,000 integrations and extensions in the Claude Marketplace; the names mentioned include Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, Docusign, HubSpot, Canva and Shopify.
• Purchases: Duet can pay via Stripe Link, using a virtual card so that the real card number is not shared.

That is a long list, and it is worth reading it soberly. These are the vendor's statements about its own product. How well this works in the day-to-day of a German mid-sized company, the page does not say, and we have not tested it.

The approval step: what the vendor writes about it

For us, the most important paragraph of the announcement is the one about approvals. According to the Claude blog, before sending a message, before a payment, before signing a document and before deleting data, Duet shows exactly what it intends to do and waits for your consent. Even with consent, the app does not trade securities or move money between accounts, according to the page. You can also disconnect any connected app again.

In addition, the page describes a safeguard for particularly sensitive information: a dedicated classifier is meant to refuse requests to disclose the strictest tier of information, including credit card numbers and identity document data. Health and personal financial information would only be shared if you expressly instruct it to.

It is noteworthy that the vendor names the limits itself: Duet can misread a thread or misjudge urgency, personal matters such as ticket bookings are said to be less mature, and the agent's browser cannot yet be viewed.

According to the page, these are the known limitations. There is also a note on costs: the background checks count toward your usage allowance, and larger tasks require more computing steps. For a business, that means: anyone who deploys a proactive agent pays not only for what it does for them, but also for the constant checking.

What users say in the announcement

The page contains three statements from users selected by Anthropic. Jess Provencher, real estate agent and owner of Pro Homes, says of a viewing appointment that had been scheduled incorrectly and that Duet had flagged: "exactly the kind of thing I want caught without having to ask". Amin Ebrahimi, owner of Shahs of Kabob with around 110 employees, used the agent for accounting, customer service and marketing and says: "I don't need to open up 100 tabs." Mya Dalton, brand strategist, says: "I love that proactivity", after Duet had flagged a client appointment shortly before her flight.

These quotes should be put in the right context. They come from the vendor's page, which publishes them, and they describe individual cases. A business with around 110 employees is similar in size to many mid-sized companies, but a restaurant business in the US works with different systems and different laws than a machine builder in Baden-Württemberg. As an impression of the mood, the quotes are useful; as evidence for your business, they are not.

Availability: not yet usable for you

We need to be clear here. According to the announcement, Claude Duet is a research preview, access runs via a waitlist, and the launch is with Pro and Max subscribers in the US. According to the page, the app is included in the Pro and Max plans and uses the same usage limits as the rest of Claude. The page names no further regions. For Germany, Austria and Switzerland, there is therefore no statement about a launch; we treat the product as not available for DACH.

• Status: research preview (test), not generally available.
• Access: waitlist.
• Region: US, initially for Pro and Max subscribers.
• DACH: no information, therefore not usable.

That means we are not describing Duet here as a tool you can use in your business today. It is a foretaste of a product class that is likely to arrive in Europe in the coming months, although we do not know a date. What matters for you, therefore, is how you prepare, so that you can deploy such an agent on your terms and not on the vendor's.

What a proactive inbox agent means in a German business

An agent that watches your inbox and apps has access to everything stored there. In a mid-sized company, that means customer correspondence, quotes, contracts, invoices, HR matters and often third-party data. This raises questions you should clarify before any deployment.

First: who has access to what? An agent connected to your account may, in the worst case, see more than an individual employee should. Define which inboxes and folders it may read at all. The ability to disconnect any connected app again is a good start, but it does not replace proper permissions set up in advance.

Second: where is the data? According to the page, Duet works on a virtual machine run by the vendor. Where it is located and how data is processed there, the page does not say, at least not in the parts we read. For you as the controller within the meaning of data protection law, this is no side issue: your customers' personal data leaves your systems. Which contracts and data processing agreements would be required for this, you would need to clarify with your data protection advisor before you start even a single test.

Third: what may the agent do without asking? The vendor has opted for an approval step before messages, payments, signatures and deletions. It is good that this step exists. Nevertheless, check for every agent you introduce: which actions count as critical? Who decides that? And what happens with actions that are not on this list, such as rescheduling appointments, creating tasks or changing contacts?

Fourth: how do you notice that the agent is wrong? The vendor itself writes that Duet can misread a thread and misjudge urgency. With a morning overview, that means an important email can get lost because the agent classifies it as unimportant. Anyone who relies completely on the summary never sees that email. A business therefore needs a way to take spot checks and to compare the overview against the real inbox.

Fifth: what does the checking cost? If background checks count toward the allowance, a one-off purchase becomes ongoing consumption. Work this through in a sample calculation before you let an agent loose on several inboxes.

Which agent at Amazing Agents is responsible: the Inbox Manager

On amazing-agents.com, the task of organizing an inbox belongs to the Inbox Manager. The website describes it as follows: it sorts customer emails by urgency and writes drafts. The homepage also states, as a general framework for all agents: every agent runs within a defined framework with approval steps for critical actions, and emails are, for example, created as drafts, not sent directly.

The comparison with Duet is instructive because it shows the core of the topic. Both approaches connect a language model with your inbox. What matters is what lies in between: which rights the agent has, what it may only prepare, and who gives approval and when. A draft sitting in your folder does no harm as long as nobody sends it. An email that has already gone out cannot be recalled. We consider this difference between preparing and executing the most important design decision for any inbox agent, regardless of the vendor.

We do not claim that the Inbox Manager is the same as Duet or can do everything Duet can do according to the vendor. They are different products with different scope. What they have in common is the pattern: the agent does the groundwork, a human gives approval.

With and without an approval step: what the difference is in the inbox

Without an approval step, an agent answers emails on its own, accepts appointments, deletes what seems unimportant and pays invoices as soon as it considers it necessary. As long as it is right, this saves time. If it is wrong, you only see it when the customer writes back, the appointment is double-booked or a deletion can no longer be undone. Errors repeat because nobody notices them. You are the sender, and you bear responsibility for what goes out in your name.

With an approval step, the agent prepares and shows you the result before anything goes out. You see the draft, the proposed appointment confirmation or the planned deletion, and you decide. This costs a moment of attention per case, but you keep the overview. As the vendor of Duet itself describes it, this includes messages, payments, signatures and deletions. For a rollout in your own business, we would start with the stricter setting and only decide, as experience grows, which parts you relax. That decision is then yours, not the agent's.

What you should prepare now

Because Duet is not yet usable for you, preparation comes before use. These three steps are useful regardless of the later vendor:

1. Write down inbox tasks. For a week, note what you actually do in your inbox: sorting, coordinating appointments, writing standard replies, forwarding invoices. This shows which tasks an agent could prepare and which should stay with you.
2. Define approval rules. Write down which actions must never happen without human consent. The vendor names messages, payments, signatures and deletions; add whatever else is critical in your business, such as commitments involving prices, deadlines or contract references.
3. Clarify data protection in advance. Talk to your data protection advisor about which inboxes an external agent could see at all, and what would be needed contractually. Anyone who clarifies this only after the test has, in case of doubt, already passed on data.

How we verified what is in this article

All information on Claude Duet comes from the announcement on the Claude blog dated October 21, 2026, which we read directly on October 22, 2026. The information on the Inbox Manager comes from amazing-agents.com. What is not in the announcement, we did not include. This applies in particular to statements on prices outside the Pro and Max plans, on data center locations and on a launch date in Europe; we found nothing on these. All feature descriptions are the vendor's statements; we have not tested the product ourselves.

The difference that matters

An agent that already knows in the morning what is on fire is tempting for many businesses. But the benefit depends less on how smart the agent is than on how clear its limits are. An agent without limits acts even when it is wrong. An agent with limits and an approval step prepares, shows you the result and waits for your okay before anything goes out in your name. That a major vendor is now building this step into its own product is a good sign. How you set it up for your business remains your task nonetheless.