In this article
Anthropic introduced Claude Duet on October 21, 2026 as an experimental app that follows tasks and works with connected applications after the user closes the screen. Initial access requires a waitlist invitation and is limited to Pro and Max subscribers in the United States. The sources checked on October 23 do not announce general availability in Brazil.
The change concerns how people allocate attention: users can delegate ongoing follow-up rather than returning to a chat for every update. That also increases the care needed around permissions, outdated information and decisions that should only happen after approval.
A cloud computer that can keep working
A conventional chatbot usually responds to a request made at that moment. A proactive agent can take responsibility for follow-up, organize steps, consult tools and report relevant developments. The difference is the work between conversations; the quality of that work still needs checking.
According to the Claude Duet announcement, the system has its own virtual machine and browser. The user’s laptop or phone does not have to remain switched on. Memory, skills and applications already connected to the Claude account provide its starting context.
This does not mean unrestricted access to the user’s personal computer. The documentation says Duet cannot see every project or local file and, during the preview, users cannot watch or take over its browser. Assessing a result depends on the materials, references and review requests the agent provides.
Work that can arrive without another prompt
The proposed workflow includes email and calendar briefings, ongoing follow-up and files produced at the end of larger assignments. The tool can consult connected information, draft replies, organize research or prepare a report. These are Anthropic’s announced capabilities; we have not conducted a performance test of Duet.
Consider a fictional freelance designer running a small business. Before a meeting, an agent could assemble authorized client history, identify unanswered questions and prepare an agenda. On finding an overdue invoice, it could draft a reminder. Sending that reminder on the company’s behalf is a separate action with a recipient and consequences to review.
Distinguishing preparation from execution makes delegation more concrete. “Look after my customers” is vague. “Prepare a summary of tomorrow’s meetings; do not send messages or change the calendar” sets a boundary that can be checked. This scenario is illustrative, not evidence that any particular service will produce the same result in every situation.
Autonomy has usage and approval limits
The getting-started guide identifies Claude Opus 5.5 as the model, with no manual model selection during the preview. Work counts toward Claude plan limits, including background activity. It is not unlimited use. Accounts are checked at intervals, so an urgent item can arrive between checks.
The guide says messages, payments, document signing and deletions require approval, and describes actions that remain off-limits even with consent. Users therefore need to examine precisely what a confirmation screen authorizes. Reviewing one draft should not become blanket permission for future messages.
Team and Enterprise are not included in the initial access described. Joining the waitlist does not guarantee an invitation date. A subscription or installed app alone does not demonstrate regional eligibility. Before buying a plan for this feature, check the current offer in the account itself.
Duet and ChatGPT Dots: announcements do not establish a winner
The official Dots guide also describes an ongoing agent with a cloud computer, memory and connected applications. The product page presents permission controls and work that continues between conversations. This puts the products in a similar broad category without proving equivalent execution.
There are concrete access differences. Duet starts in the United States with Pro and Max invitations. Dots documentation describes a gradual rollout to adult Pro users in eligible markets, with regional exclusions, and access arrangements for certain business environments. A Brazilian account should not infer Duet access from a ChatGPT offer, or assume the reverse.
For a useful comparison, ask operational questions. Which apps can this account actually connect? Can the agent use its own browser? Can the user observe the work? Which actions need confirmation? How can work be paused, apps disconnected and usage reviewed? Answers depend on the service and rollout stage; an announced capability does not replace these checks.
IANautaLab has not run a comparative benchmark. This is an analysis of documentation. A business’s choice depends on its workflow and privacy requirements as well as on the quality of generated text.
Disconnecting does not erase everything already read
Agents connected to email and documents encounter material written by other people. A message can contain malicious instructions intended to redirect the task, a risk known as prompt injection. More ordinary mistakes also matter: confusing customers with similar names, treating an old negotiation as current or flagging an appointment that has already been cancelled.
Start with a narrow goal and the smallest useful set of connected applications. Duet’s documentation says disconnecting an app stops future access, but memory of information previously obtained can remain. Revoking access and deleting information already incorporated are different operations.
IANautaLab’s AI in everyday life section provides context for this kind of delegation. In practice, begin with preparation that has no external impact, inspect the sources used and expand the scope only after understanding the controls. Time savings can only be assessed after counting the review work too.
Sources consulted
Anthropic — Claude Duet announcement. Claude Help Center — access, limits and privacy. OpenAI — Dots guide. ChatGPT — Dots overview. Provider documentation checked on October 23, 2026; this is not an independent product test.
