Anthropic's Claude Duet Will Now Check Your Email Every Few Hours So You Don't Have To Worry About It Yourself

For years, the knock on AI assistants was that they sat there. You had to talk to them. Like a houseplant with a context window, they waited patiently until you remembered they existed.

On October 21, Anthropic announced the cure for this: Claude Duet, an assistant that does not wait. It checks your connected apps every few hours, assembles a morning brief, and messages you when something needs attention. It is, in the company’s framing, a “proactive agent.” In the rest of the world’s framing, it is a coworker who has finally learned to say “just circling back.”

The product is a research preview, built around Claude Opus 5.5, and available by waitlist to US Pro and Max subscribers. It lives in its own app on iOS and Android and on claude.ai, separate from the main Claude app, in the way a vacation home is separate from your regular home: same owner, more maintenance.

Duet runs tasks in the background on its own virtual machine, which means your laptop and phone do not need to stay on. This is a genuine breakthrough in the history of work. For the first time, a task can be in progress without anyone’s lid being open. Somewhere in a data center, a small computer you will never meet is reconciling your invoices, and it will not ask for a recommendation on LinkedIn afterward.

According to The Neuron’s coverage, the sorts of things it prepares include morning briefings, draft follow-up messages, invoice reminders, and payroll worksheets, drawing on a reported 6,000-plus connectors. That is a lot of doors to have keys to. Anthropic’s answer is the approval gate: before Duet sends a message, makes a payment, signs a document, or deletes anything, it “shows you exactly what it’s about to do and waits for your approval.”

Waits for your approval. Reader, consider the emotional arc of this. The agent does the work. You do the checking. The agent is doing the part of the job that was fun, and you have been promoted to the part where you read a draft payroll worksheet at 7:40 a.m. and try to decide whether it feels right.

To be clear, the safeguards are not a joke. Some actions are off-limits even with approval, such as trading securities or moving money between accounts. Duet’s credential handling follows the design used for Claude Managed Agents, so it “never directly sees your passwords or sign-in tokens,” which is the correct amount of access to have to your passwords, and also, coincidentally, the amount most coworkers have after the third reset email. A dedicated classifier is meant to block leaks of the strictest-tier data, like credit card numbers and government ID. Anthropic says red teams and an external firm penetration-tested the thing, and that it is covered by a public bug bounty.

What Anthropic does not claim is perfection, which is refreshing in a product category where “it just works” has historically meant “it worked once, in the video.” The company’s own list of limitations says Duet may misread a thread or misjudge urgency, that personal errands such as booking tickets are weaker than work tasks, and that it cannot yet show you its browser. So the assistant that handles your business correspondence cannot yet show you what it was looking at. Think of it as a very diligent intern who works behind a frosted-glass door.

The Neuron’s analysis supplies the question that the launch copy politely steps around: does the time saved exceed the time spent checking? The review burden is real. If you have to reopen every app to verify what Duet found, the savings shrink. Payroll is the sharp example, because an omitted amount can look exactly like a complete worksheet. A spreadsheet that is missing a line item does not announce itself. It simply sits there, looking finished, like a soufflé that has not yet disclosed what it has done.

There is also the cadence problem. Duet checks every few hours, not continuously, so a last-minute cancellation could sit unnoticed until the next sweep. “Proactive,” in this context, means “proactive on a schedule,” which is how most of us were proactive in our early careers, right up until the meeting.

And the plumbing: background checks and larger tasks count against your plan’s usage limits, and Anthropic has not published what a typical workday costs in usage. So the assistant that saves you time is also quietly spending your allowance while you sleep. Somewhere a meter is running, and it is a polite meter.

Competition is already circling the idea. The Neuron notes that Meta’s Muse for Small Business and OpenAI’s Dots also monitor services and request approval before acting, and MacRumors frames Duet as Anthropic’s entry into that same proactive-agent race. The whole industry has arrived at the same destination from three directions, which is the product-management term for “everyone read the same memo.”

Perhaps that is the real story. The chatbot era taught us to ask. The agent era is teaching us to approve, a verb that is rapidly becoming the central activity of white-collar life. Soon the working day will consist of a gentle stream of notifications reading “Claude would like to send this,” and your job will be to look at it for about four seconds and tap the green button. It is the most efficient possible way to remain responsible for everything.

Duet’s waitlist is open to US Pro and Max subscribers. Team and Enterprise plans are not included in the preview. Anthropic says Duet asks before it acts. Whether you ask before you approve is, as ever, between you and your payroll.

Sources: