Claude Duet’s real test is whether AI can leave you in control

What if the next AI breakthrough is not a better answer, but a better interruption? That is the wager behind Claude Duet, Anthropic’s research preview for an assistant that watches connected apps, decides what deserves attention, and keeps working after you close the app. The product arrives as Silicon Valley’s “always-on” agent race accelerates—and it forces a question the industry would rather leave vague: when does helpful initiative become another layer of work to supervise?

Anthropic says Duet sends a short morning brief, checks connected accounts every few hours, and contacts users only when something needs attention. It can run tasks in a cloud virtual machine with its own browser, using an ongoing conversation rather than separate chats. The product is available through a US waitlist on Pro and Max plans, according to Shattered’s account of the launch. The launch has also drawn attention in online discussions.

Duet’s innovation is judgment, not autonomy

The important distinction is not that Duet can execute tasks. Claude already has memory, skills, connected apps, browsing, and scheduled work through Cowork. Duet’s new behavior is initiative: it decides that something in your digital life may deserve your attention before you assign a task or create a schedule.

That sounds subtle, but it changes the product relationship. A chatbot waits for a question. A conventional automation waits for a rule. Duet occupies the uncomfortable middle: it interprets an evolving stream of calendars, messages, invoices, and documents, then chooses when to interrupt a person.

That is potentially valuable for professionals and small-business owners whose work is fragmented across tools. A calendar conflict, an unpaid invoice, or a message that needs a reply can be surfaced without requiring the user to remember which system contains it. Anthropic’s examples include scheduling, dashboards, forecasts, prototypes, and web apps.

But we think the interruption policy is more important than the task catalog. An agent that notices everything is not useful; it is a new notification system with a persuasive writing style. Anthropic acknowledges that Duet can “misread a thread or miscalibrate what’s urgent,” and says it is tuning that behavior using tester feedback. That admission is not a footnote. It is the central product risk.

Approval gates protect actions, not attention

Duet asks for approval before sending a message, making a payment, signing a document, or deleting anything. It cannot trade securities or move money between accounts, even with approval. Users can disconnect an app, and Anthropic says the agent does not directly see passwords or sign-in tokens.

These are sensible boundaries. They also reveal the limit of the industry’s current safety model: it is easier to identify a consequential click than to identify a consequential interpretation. A mistaken payment request is visible. A bad judgment about which customer issue is urgent may be less dramatic, yet it can still distort a workday.

The distinction between reading and writing is especially revealing. Duet can check QuickBooks for unpaid invoices without approval, but drafting reminder messages stops at the approval stage because messaging another person is an external action. That is a reasonable line, but it leaves users responsible for deciding whether the agent’s summary, context, and proposed wording are correct.

Anthropic says a classifier approves actions, defenses address hidden instructions and prompt injection, and a dedicated classifier checks outgoing messages for sensitive information. Those safeguards matter, particularly when an agent crosses multiple services. Still, no classifier can turn delegated judgment into personal judgment. Approval can become a rubber stamp if the system presents too many small decisions or explains them poorly.

Opus 5.5 makes Duet’s economics part of the design

Duet is built around Opus 5.5, Anthropic’s top general model in the supplied launch material. That choice makes sense: the product is not merely generating text; it is deciding what can wait, what requires a person, and which sequence of tools might complete a task.

It also creates a cost question. Background checking, browsing, and multi-step execution consume capacity even when the user is not actively chatting. Anthropic’s pricing page says usage limits apply across Claude on web, desktop, mobile, and Claude Code, with activity depending on conversation length, model, and features. In other words, “always on” does not mean unlimited. It means a larger share of the user’s allowance may be spent on work happening outside the visible conversation.

That could be worthwhile if Duet reliably eliminates coordination overhead. It could feel wasteful if the result is a stream of low-value alerts, unnecessary checks, and approval requests. The product therefore needs a clearer measure than model quality: useful attention recovered per unit of AI usage.

The winning assistant should expose its reasoning about work

The strongest part of Duet’s design is the ongoing thread. Quick replies and longer projects live together, while memory, skills, and connected apps carry over from the Claude account. That continuity could make the assistant feel less like a collection of isolated prompts and more like a shared workspace.

But continuity can also hide complexity. When an agent quietly checks several accounts, chooses tools, and returns with a result, users need a legible record of what happened: which sources it consulted, what assumptions it made, what it changed, and what remains uncertain. The approval dialog should not merely ask “Proceed?” It should make the scope of the decision understandable.

This is where Claude Duet becomes a meaningful test for AI products generally. The question is not whether agents can perform more steps without supervision. The question is whether they can take on coordination while leaving people with a clear mental model of the work.

Our verdict is cautiously favorable—but only for users willing to begin with narrow permissions, limited notification scopes, and reversible tasks. Duet looks more consequential than an upgraded chat window because it introduces initiative. That is also why it deserves skepticism. If Anthropic can make its interruptions sparse, explainable, and genuinely useful, Duet may advance the collaborative-assistant idea. If not, it will simply automate the burden of managing an AI coworker.