Claude Duet: what this AI agent does and what businesses in Spain can do now

Claude Duet is Anthropic's proactive AI agent app: it reviews accounts you connect, identifies things that need attention and prepares work without a separate request for every step. Announced on October 21, 2026, it runs on Opus 5.5 with its own cloud computer, according to the official announcement and product page. For now, it is a research preview with a waiting list for Pro and Max users in the United States; Spain is outside this phase, as the getting started guide explains.

Published on October 29. Sources consulted up to October 23 and checked again on October 29, with no changes to what this article covers. This article draws on documentation and other people's early experience. Our proposed Soul IA evaluation below describes how we would assess it; it does not claim that we have tested Duet.

What proactive actually means here

The idea is that some work starts without a specific instruction: a morning briefing, suggested priorities or drafts of next steps. Duet can also keep working while you are away. Our reading of Anthropic's description is that you are delegating some of the judgement about what deserves attention, as well as the work that follows. That judgement is what we would want to assess.

Two examples in the launch announcement make this concrete. For Jess Provencher at Pro Homes, Duet spotted a property viewing scheduled on the wrong day without being asked to check it. For Mya Dalton, it prepared a meeting and suggested moving it after finding a flight conflict. Anthropic selected these testimonials. They explain the intended use, but do not establish a general success rate.

Duet checks connected accounts every few hours, according to the announcement. If you need urgent alerts, allow for gaps between checks. Finding something, drafting a response and taking an external action are separate stages. The announced design requires approval before sending messages, making payments, signing or deleting.

The help guide lists Gmail, Google Calendar, Google Drive, Slack, QuickBooks, HubSpot, Docusign and Calendly. The product page advertises more than 6,000 connectors and plugins combined. That total does not tell you whether each integration supports the same operations. For a business process, we would check exactly what the relevant connection can read and write.

The limits of the current preview

• Microsoft 365 is not available yet. The Duet guide states this restriction, even though its logo appears on the product page. Do not plan an Outlook workflow around that logo.
• There is no model selector. The product page names Opus 5.5, and the help documentation confirms that you cannot choose another model.
• Duet cannot watch or take over your browser, call third parties or send them SMS messages. It cannot see local computer files or all your projects. It can read Claude work but cannot modify it. These are limits in the current guide.

The announcement describes 1Password and an expanded visual Shopify editor as future additions. We have left them out of the available features. The research also could not confirm native Windows or macOS apps, full conversational voice or specific Spanish language coverage, so we make no promises about those.

What early users report

Dan Shipper at Every describes two weeks of early access. He found value in the suggestions, channels and task tracking. His reservations include email draft quality, the absence of conversational voice and scheduled runs that did not reach his custom connectors. Every says Anthropic supplied access without influencing the review. These observations describe his particular setup and use.

Claire Vo in Lenny’s Newsletter tried email, Slack, a sales funnel and content creation. She liked the visual organisation, but reported delays, less initiative than she expected and a lack of collaboration between multiple users. Her criticism of the missing desktop app should be read alongside the officially announced web access.

Alex Volkov on ThursdAI showed the web interface on October 22 and liked the context Duet already had. He had just set it up and acknowledged that his testing time was very limited. The link goes to the programme, where the content can change. We treat this as an initial impression, without drawing conclusions about sustained reliability.

These accounts offer useful questions to investigate. They do not establish your likely savings. Carl Franzen at VentureBeat points out the absence of published productivity and cost per task metrics. The research gives us no basis for a general accuracy percentage, a promise of hours saved or a financial return.

Availability and pricing for a business in Spain

The announced interfaces are standalone iOS and Android apps and web access through claude.ai. The product page and guide place the preview in the United States, with a waiting list for Pro and Max. Team and Enterprise are not included yet. The sources reviewed give no confirmed date for Spain or expected invitation time.

Duet is included during the preview, but shares usage limits with the rest of Claude, including its background work. The help guide describes five hour windows and weekly limits. We would therefore measure what its checks and preparation consume even while you are away from the keyboard. Inclusion in a subscription does not tell you how much useful work the available quota will cover.

The published prices are $20 per month or $200 annually for Pro, and Max from $100 per month. These are dollar references, without treating them as final Spanish prices. The reviewed sources publish no fixed price per task and confirm no price for Duet after the preview.

Permissions and security before connecting accounts

Anthropic says credentials are handled separately so that the model does not see passwords or sign in tokens. Its announcement also describes sensitive information filters and protections against malicious instructions. These are claims about the design; the research does not independently establish their effectiveness. Anthropic's trustworthy agents framework acknowledges risks from misunderstood intent and instructions hidden in external content.

The general connector documentation says access inherits the user's permissions in the original service. We would distinguish that scope from approval for a particular action. Disconnecting an app ends future access, but memory from earlier conversations can remain, according to the Duet guide.

Pro and Max fall under the consumer training policy. A paid subscription does not rule out all training use: permissions, consent and certain safety reviews matter. Raw connector content is excluded from that training set, but can enter it when copied into a conversation. Deleting a chat also does not necessarily delete memories it generated; memory has separate controls.

What a Spanish small business can do meanwhile

Start with one specific need and an option covered by the research. The table sets out scope and conditions. It does not claim that these products reproduce Duet or that every account has access to every gradual rollout. Choose the process before choosing the tool.

A daily Cowork briefing using approved sources and preparing drafts would be a routine you define. Its scheduled tasks can use connectors and run without a connected device, except where local dependencies are involved. It would be inaccurate to say Cowork always needs your laptop on: new Pro and Max tasks have run in the cloud since October 6; older local tasks remain local, according to the documentation.

Claude Tag was announced on June 23, 2026, in beta for Team and Enterprise. It works in Slack and can offer updates when ambient behaviour is enabled, as described in its announcement. Connecting Slack to Duet and installing Tag involve different contexts and permissions. With n8n, a Schedule Trigger runs your workflow at specified times or intervals. You supply the decision logic.

How we would assess it at Soul IA: one process for one week

We would choose one measurable process for a week. For example, preparing a morning list of outstanding items from sources you authorise. Before starting, we would write down what belongs in the list, what is out of scope and which actions need approval. That gives us a reference for judging the output, rather than relying on how impressive a suggestion feels.

1. Correct results: which outputs meet the agreed criteria and are usable for the next step.
2. Omissions: which relevant items the process should have found but missed.
3. Corrections: what you had to change and how much effort the review took.
4. Usage: how much quota preparation and background work consume.

At the end, we would decide whether to expand the scope, adjust the routine or stop. This is a proposed evaluation method, with no claim of Soul IA test results or customer cases involving Duet. For related reading on usage, see our article about tokens:

How to spend fewer AI tokens in Claude Code →

Frequently asked questions

Bring it back to one process in your business

We can talk through the first process: the result you need, who would review it and how to measure whether it helps. That gives you a concrete trial to consider with available tools, without building a plan around an unconfirmed Duet launch date.

Talk through a specific process →