Claude Duet: what the first week of reviews found

Anthropic opened Claude Duet as a research preview the week of October 21, and the search behavior followed immediately: people are looking for the waitlist, the access rules, and reviews from anyone who got in. Duet is a phone-first agent that messages you like a person, works on its own computer in the cloud after you close the app, and asks before it sends a message, makes a payment, signs a document, or deletes anything. Three reviewers published real first-hand accounts in the launch week. Read together, they converge on one lesson that matters well beyond this product: an agent is only as useful as the loop that brings you back to a decision.

Access, cost, and the fine print

Per Anthropic’s help page, checked October 30, 2026: Duet requires a Pro or Max plan, a US account, and a waitlist invitation that goes out in waves with no promised date. One reviewer on the top Max tier waited eight days. Team and Enterprise plans are out for now, and the invitation email must match both your Claude account and the Google or Apple account on your phone.

There is no extra charge during the preview, but it is not free in the way that matters: every background check, morning brief and quiet monitoring pass draws on the same five-hour and weekly usage limits as your chats. Anthropic deliberately checks accounts only every few hours because constant checking would drain plans faster. New admits get one full limit reset, claimable until November 30, 2026.

The hard limits from the same page: it cannot see the files on your computer, it makes no phone calls or texts to other people, Microsoft 365 is not available yet, you cannot watch or take over its browser, and some actions - trading securities, moving money between accounts - are blocked even if you approve them. Email sends from your connected accounts require your approval each time.

What three early reviewers found

The starting point is genuinely good. Two of the three praised the onboarding, and the reason is structural: Duet inherits the connectors, memory and skills already on your Claude account, so it starts useful instead of empty. One reviewer with sixteen Claude projects asked it to find half-finished work and got back eight concrete items with offers to help - something no fresh-start agent could have done.

Follow-through is the open question, and the reviews diverge. Claire Vo’s launch-day review (October 21, tested on early preview builds) is the cautionary one: she handed Duet ticket shopping and document work, tasks ran ten minutes or more, and the follow-up was quiet enough that she forgot the open threads entirely. Her point is the one to keep: watching a task list fill with activity is not the same thing as being brought back to make a decision. Gabriel Elizondo’s first day (October 28) showed the opposite pattern - Duet messaged him unprompted about a developing story in his beat, flagged how far the source could be trusted, kept checking, and came back in the evening to say the claim had not held up. Same product, same week, opposite experiences of the feature that matters most. Chris Alarcon’s day one (October 29) landed in the middle: Duet produced a clean summary of an insurance chase his other agent was already running - pleasant, and it moved nothing forward.

Speed splits in two. The interface feels fast; the work behind it is slow. Vo’s point is that this ordering matters more than it seems: a slightly less capable agent that keeps you in rhythm beats a stronger one you wait on, because you stay engaged. Elizondo noticed the same lag and shrugged it off as launch-week infrastructure.

The walls are real. Duet runs on Anthropic’s computer, not yours - Alarcon hit that immediately when he pointed it at his own machines. Vo, testing days earlier, had no desktop experience at all, which broke the workflow she actually wanted: start on the phone, finish at her desk. And it is single-player today; channels organize one person’s work but nothing is shared with the five humans a small business owner typically works with.

What to do before you connect an agent like this

The reviews double as a setup checklist, whatever agent you are evaluating:

• Audit the connector list first. Inherited connections go live immediately and the agent starts checking them within hours. Look at what you connected months ago before something starts reading it every few hours.
• Test the return trip, not the request. Anyone can make an agent accept work. Start something async, then confirm you can review the output, answer its questions and approve the next step from where you actually work - including whether it comes and gets you.
• Compare its priorities against yours. Let it triage your inbox, then judge its ordering against your own before letting it sequence your day. Finding activity is easy; knowing what deserves attention is the hard part.
• Price the background work. Monitoring is not free just because it is quiet. Check what briefs and check-ins cost against your plan’s limits, and claim the one free reset before November 30 if you are in the preview.
• Know what persists. Disconnecting an app stops new access; what the agent already learned stays in its memory until you reset it.

Why this matters for outreach specifically

Every reviewer, even the most positive, described human judgment appearing at every consequential point - which priorities were right, whether a draft met the standard, whether to act. That is not a limitation of one product; it is the shape of the category. And it has a sharp edge for outreach: the approval gate is only as good as the loop around it. An agent that drafts follow-ups but never brings you back to approve the send leaves replies aging in the Other tab and sequences stalled mid-conversation. Silence after the draft is the same failure as silence before it, and silence is hard to read - as we covered in where your LinkedIn message actually goes before they accept.

So when you evaluate any outreach agent, do not ask how elegantly it accepts work. Ask whether it returns with decisions ready and makes your part fast. The send stays yours - the loop has to close.

Dollar Outreach is built on exactly that principle: you paste profile URLs, it queues the note, the invitation and the follow-ups under fixed daily caps, and you read and send every message yourself. Nothing goes out without your click, and nothing waits on a machine deciding you have seen it.

Limits

• Access, pricing and feature facts from Anthropic’s “Get started with Claude Duet” help page, checked October 30, 2026. This is a research preview; limits and availability can change without notice.
• Reviewer experiences are self-reported from builds days old at test time (October 21-29, 2026). The follow-through divergence above may be settings, notification configuration, or the product changing under them.
• One reviewer tested before a desktop experience existed; another’s account suggests the surface area is already expanding. Treat every specific here as perishable.