Claude Duet works between messages. Its background checks also spend your allowance

Claude Duet is Anthropic’s attempt to move the assistant’s starting point from an empty chat box to the work already accumulating in your inbox and calendar. Announced on October 21, the research preview checks connected services, prepares work in the background and asks for decisions when it reaches an action that needs approval.

The immediate catch is access. The official product page offers a waitlist for Pro and Max subscribers in the United States. Team and Enterprise plans are excluded for now. This is neither a worldwide launch nor a separate unlimited allowance: background checks and larger tasks consume the same plan limits as the rest of Claude.

That last detail changes the proposition. An assistant can use resources while you are doing something else. The question is whether the work it brings back saves more attention than its requests, mistakes and usage costs consume.

Initiative is the new feature

According to Anthropic’s launch explanation, Duet runs on Opus 5.5 and works through its own cloud computer. It can keep going when the user’s device is off. Messages and voice requests share an ongoing thread, with the account’s existing memory, skills and connections available to the agent.

The proposed rhythm is a morning briefing followed by background checks every few hours. That is different from watching every incoming event continuously. Something urgent can arrive between checks, and Anthropic acknowledges that the preview may misread a conversation or judge its urgency badly.

For a small business, finding an unpaid invoice and preparing a reminder are plausible examples of useful preparation. They are also examples where context matters: a late payment might already be the subject of a phone agreement that never reached the connected apps. The assistant’s ability to collect records does not establish that those records tell the complete story.

Approval is a boundary, not a quality guarantee

Anthropic says Duet can read, draft and prepare work, but must show the proposed action and wait before operations such as sending messages, paying, signing or deleting. Securities trading and money transfers remain outside its permitted tasks even with approval. These are the company’s stated operating boundaries, rather than results of our testing.

An approval prompt is useful only if the human can understand the decision. A draft reminder should make its recipient, amount and wording easy to inspect. A proposed purchase should expose the merchant and terms. Otherwise the work has simply moved from performing the task to reconstructing what the assistant intends to do.

Anthropic’s broader research on trustworthy agents treats reliability as a property of several layers: the model, its operating framework, tools and environment. That is a helpful way to understand the risks here. A convincing response is not sufficient when the system also reads outside material and can act through connected accounts.

An email, web page or document can contain instructions that conflict with the user’s intentions. Anthropic describes defenses against such prompt injection, but the meaningful question remains what access is available and what checks stand between a mistaken interpretation and a consequential action.

A preview with a narrow launch territory

Duet is offered through standalone iOS and Android apps and the web. The initial US restriction matters to international readers: the announcement does not establish an Italian or wider European release date. A normal Claude subscription in another market should not be mistaken for eligibility for this preview.

The product also cannot yet show the user its browser as it works, according to the launch notes. That limitation makes clear summaries and reviewable outputs particularly important. Our earlier coverage of agents exposing the steps behind their answers addresses a related problem: a finished result and an inspectable process offer different kinds of confidence.

Duet’s useful ambition is to reduce the effort of noticing what needs doing. Its hardest test is more selective: recognizing when an apparently routine task contains a decision that belongs to the person. A good morning brief should leave its reader with fewer loose ends, not a new job supervising the brief.