A chatbot that waits for a question is one thing. An assistant that watches connected work apps, notices a loose end and prepares the next step is another. Anthropic's newly announced Claude Duet is aimed at that second job, but its launch comes with limits that matter as much as the demonstration.
Announced on October 21, 2026, Duet is a research preview, initially offered through a waitlist for Pro and Max subscribers in the United States. It is not a general worldwide release. AsiaNews Iran examines what that means for access, control and everyday use.
What is Claude Duet supposed to do?
Anthropic describes a separate app for iOS and Android, also available through Claude's website. Rather than starting a new chat for each task, users keep a continuing conversation. Connected work services give it context for briefings and follow-up work.
Reuters reported that the product targets professionals and small-business owners, with examples including preparing meetings, forecasting finances and creating web applications. These are advertised uses, not evidence that every task can be delegated reliably.
The distinction is practical. Preparing a meeting brief can save time even when someone checks it. Quietly misunderstanding an email thread, however, could send that person into a meeting with the wrong priorities. The useful question is not simply whether an agent can produce work, but how easily its work can be checked.
Who can join, and what uses the allowance?
According to the launch announcement, access is included in the eligible subscriptions rather than sold as a separate Duet plan. A subscription does not itself remove the waitlist. Readers outside the initial US rollout should not mistake the announcement for confirmation that their account can use it.
MacRumors reports that Duet works in a cloud-based virtual machine, so the user's computer need not remain switched on. Background activity still counts toward the subscription's shared usage limits. An assistant doing work out of sight is not an assistant working without cost to the allowance.
That makes a small trial more informative than connecting an entire working life at once. Start with a clearly bounded task, review the output and watch how much capacity it consumes. A tool that helps with tomorrow's meeting is less useful if today's experiments exhaust the allowance needed for other work.
Where the human approval boundary sits
Anthropic says Duet asks before consequential steps such as sending messages, making purchases, signing documents or deleting files. Its announcement also excludes activities such as securities trading and money transfers.
Preparation and execution should therefore be treated as different stages. A proposed email needs its recipient, attachment and wording checked, not merely a glance at a polished summary. The same principle applies when a task crosses between personal and workplace accounts.
Access to an app also raises a separate question from permission to act. A user should know which account is connected and whether the material belongs in that service at all. Our earlier report on DeepSeek and federal agency restrictions illustrates why organizations may draw different boundaries around AI tools. It does not establish that the same restrictions apply to Duet.
A preview, not a substitute for judgment
The company acknowledges that Duet can misread a conversation or get urgency wrong. Launch coverage establishes that the product has been announced; it is not an independent guarantee of accuracy or security.
For early users, the sensible test is modest: does it surface something genuinely useful, prepare it correctly and leave the final decision clear? If checking the result takes longer than doing the task, the impressive part of the demo has not yet become useful assistance.