Most AI assistants wait for a question. Claude Duet is designed to notice important changes across connected apps, prepare work in the background and ask for approval before consequential actions. Here is how Anthropic’s new agent works, what it can do today and what still needs testing.
Quick Take
- What launched: Anthropic opened a Claude Duet research preview on 21 October 2026. Access starts through a waitlist for US Claude Pro and Max subscribers; this is an early product, not a worldwide release.
- What changes: The agent can assemble a morning brief, check connected apps periodically, flag items that may need attention and continue delegated tasks after you step away.
- Where people approve: Anthropic says Duet requires approval before sending messages, paying, signing documents or deleting items; certain financial actions remain prohibited.
- What is unproven: Anthropic acknowledges that Duet can misread a thread or misjudge urgency. There is no independent evidence yet that it reliably manages an entire workday across every organization.
What Is Claude Duet?
Anthropic describes Claude Duet as a separate app for iOS, Android and the web, built around Claude Opus 5.5. It draws on a user’s Claude memory, skills and authorized connected apps. You can still ask it for work in a conversation, but it can also initiate contact: a daily brief may highlight a meeting conflict, an overdue document or a message that needs a reply.
That makes proactivity the defining idea. The valuable result is not a notification for every email. It is an alert that correctly identifies a decision worth your time, cites the underlying information and lets you move from insight to an appropriate next step. If the agent cannot reliably distinguish urgent work from ordinary noise, its ability to message first could become a distraction.
The timing matters. This is a research preview, initially offered through a US waitlist to Pro and Max subscribers. Anthropic has not presented it as a generally available enterprise service or announced a Sweden-wide rollout in its launch article.
How Does Claude Duet Work While You Are Away?
Anthropic says Duet runs in a separate virtual machine with its own browser. A delegated task can therefore continue when your laptop is shut, and the app can notify you when work is ready or a decision is needed. Duet keeps one conversation thread for typed or spoken requests, while the underlying job can move among permitted services.
Its attention is scheduled rather than uninterrupted. Anthropic says it checks connected accounts every few hours; the frequency can be adjusted, and the checks consume the same plan allowance as other Claude work. That makes the phrase “always on” misleading if readers take it to mean live monitoring of every event.
The workflow is best understood as a sequence: gather permitted context → decide what matters → draft or execute a delegated task → request approval where required → report the result. Each transition is a possible failure point. An outdated calendar entry can produce a mistaken alert; a misleading webpage can try to steer the browser; a correct draft can still be sent to the wrong recipient if review is careless.
Claude Duet Versus a Regular Claude Chat
The same model family may power both experiences, but the product behavior is different. The table compares the workflow a reader should expect, not a benchmark of model intelligence.
| Question | Regular AI chat | Claude Duet research preview |
|---|---|---|
| Who starts? | You open a conversation and provide a request. | You can delegate a task, and Duet can also send a brief or alert you from connected app activity. |
| Where does context come from? | Primarily your prompt and any files or tools available to that chat. | Permitted Claude memory, skills and connected apps support continuing work across tasks. |
| Can work continue after you leave? | It depends on the particular product workflow. | Anthropic says Duet uses a cloud virtual machine and browser for background tasks. |
| Who authorizes sensitive actions? | It depends on the tools and settings of that chat. | Anthropic says sending, paying, signing and deleting require explicit approval. |
| What determines usefulness? | Answer quality to the prompt you supplied. | Alert relevance, task completion, correct permissions and manageable review time. |
What Can a Proactive AI Agent Actually Do at Work?
Anthropic says Duet can schedule meetings, assemble dashboards, prepare financial or sales forecasts, prototype web apps and work with services available through the Claude Marketplace. Those are product capabilities described by Anthropic, not a promise that every account has identical integrations or that the result is correct without checking it.
Summarize permitted email, chat and calendar items, identify a conflicting appointment and propose a revised time. The user verifies the people, dates and source messages.
Use approved CRM history to draft an account update, find missing information and prepare a follow-up email. A person reviews the facts and approves the recipient and final text.
Check authorized invoice records, prepare a reminder or generate a forecast. A finance owner confirms amounts and assumptions before any customer-facing or financial action.
Gather the latest status from connected tools, build a draft dashboard or prototype an internal app. The team tests the output and controls publishing or deployment.
A connector catalog is a menu of possible paths into services. It is not proof that a connected workflow has been configured with the right scope, consent, data quality or failure recovery. A responsible team starts with narrow read permissions and draft outputs, then expands only after testing.
Who Is in Control When Claude Duet Takes Action?
Anthropic says Duet can read, draft and prepare work independently, but it shows a proposed action and waits for approval before sending a message, making a payment, signing a document or deleting information. Trading securities and moving money between accounts remain off-limits even if a user approves. People can disconnect connected apps.
The company also describes controls for sensitive data, credential handling, prompt injection and action selection. Its announcement says a dedicated classifier is intended to reject disclosure of highly sensitive identifiers and that sending certain other sensitive information requires explicit instruction. These are descriptions of Anthropic’s safeguards, not a guarantee that an agent cannot make a mistake.
Prompt injection is a practical concern for any browsing agent. A page or message can contain instructions that look like part of the job but actually attempt to redirect the agent. Teams should keep the source of every proposed action visible, log what the agent accessed and test whether an untrusted document can make it change recipients or disclose private content.
What Does Claude Duet Cost in Time, Attention and Usage?
During the research preview, Anthropic says Duet is included with eligible Pro and Max plans and shares their usage limits. Background account checks count against those limits; a large task such as building a forecast or a deck can use more model turns than a short chat. There is no reason to assume that leaving it active produces unlimited work at a fixed practical cost.
Human review also has a cost. A useful agent saves time only if the brief is accurate, alerts are selective and approval screens make consequences clear. Count the minutes spent verifying drafts and correcting errors alongside the minutes the agent spends running tools. Anthropic says the current version may misread a thread, miscalibrate importance and perform less strongly on personal errands. It cannot yet show users its browser during a task, according to the launch article.
How Should a Team Test It Before Trusting It?
- Begin with a limited dataset: Grant access only to the apps and folders needed for one clearly defined workflow. Exclude personal and sensitive material unless the task genuinely requires it.
- Measure alert quality: Over a fixed pilot period, count useful alerts, missed urgent items and unnecessary notifications. Check whether each conclusion links back to current source data.
- Rehearse approvals: Ask the agent to prepare an email, a calendar change and a document action. Verify that recipients, amounts, attachments and permissions appear before a person approves.
- Test adversarial content: Place misleading instructions in a document or webpage that the agent must summarize. Confirm that it treats those instructions as untrusted source content.
- Calculate total cost: Record plan usage, elapsed task time and human correction time per accepted outcome. Compare the result with doing the same work through your current tools.
The metrics should determine the next stage. A single impressive demonstration does not establish dependable performance across months of changing colleagues, software permissions and business rules.
What About Privacy, Copyright and the EU AI Act?
Duet’s US-first research preview does not establish availability in Sweden. An organization planning future EU use should map which personal data the agent can read, the purpose and legal basis for that processing, retention, third-party recipients and employee access. The European Commission’s GDPR guidance emphasizes purpose limitation, data minimization, security and accountability.
The European Commission’s AI Act guidance explains that obligations depend on the actor’s role and the specific AI system or use. A general workday briefing should not automatically be called high-risk; a system used to make or support certain employment or credit decisions may raise different requirements. Human approval at a button does not by itself settle the legal classification or remove other obligations.
For generated drafts, dashboards and campaigns, check the rights to source documents, images and customer material before publishing. This article is original editorial writing; links are included so readers can inspect the product announcement and EU guidance themselves.
MaGeN-AI View
The real test is judgment
Claude Duet offers an interesting change in AI interaction: the assistant may bring useful work to you instead of waiting inside a chat window. Its strongest early use is likely to be finding relevant work, preparing a verifiable draft and asking the right person at the right moment. The research preview still needs to demonstrate that it can do this consistently, without flooding users with alerts, misreading private context or turning an easy task into a long approval process.
Frequently Asked Questions
What is Claude Duet?
Claude Duet is Anthropic’s proactive AI agent app. It can prepare a daily brief from connected apps, alert users to relevant changes, and carry out delegated work in the background.
Is Claude Duet available in Sweden or the EU?
Anthropic’s October 2026 research preview begins by waitlist for US Claude Pro and Max subscribers. The company has not announced general availability in Sweden or across the EU in its launch article.
Can Claude Duet send emails or make payments without approval?
Anthropic says Duet shows the proposed action and waits for approval before sending messages, making payments, signing documents or deleting items. It says securities trading and transfers between accounts remain off-limits even with approval.
Does Claude Duet work when my computer is off?
Anthropic says the agent uses a separate virtual machine and browser, so delegated tasks can continue without a user’s laptop or phone staying on.
Does Claude Duet have separate usage limits or enterprise availability?
During the research preview, background checks and larger tasks consume the same usage allowance as the subscriber’s Claude plan. Anthropic’s announcement describes US Pro and Max waitlist access, not a broad enterprise rollout.
