Anthropic launches Claude Duet, an agent app that works in the background
On October 21, 2026 Anthropic introduced Claude Duet, a research preview built for professionals and small business owners. It assembles a morning brief, watches connected apps while you are away and asks before consequential actions. Access starts by waitlist for Pro and Max subscribers in the US.
Anthropic opens an agent app that lives outside the chat window
Anthropic's announcement post is dated October 21, 2026 and presents Claude Duet as a research preview. TechCrunch covered the app a day later, on October 22. According to Anthropic, Duet surfaces items that need attention, finishes the tasks it is given, keeps working while the user is away, and checks with the user before acting on their behalf.
The routine has two parts. Each morning Duet gathers the most important items across connected apps into a daily summary. During the day it checks those apps every few hours and sends a message only when something needs attention; users can adjust what they hear about and how often.
Its own virtual machine, its own browser and one continuous thread
Duet runs in its own virtual machine and its own browser, so a laptop or phone does not have to stay on. Users talk to it by typing or by voice in a single ongoing conversation, and memory, skills and connected apps from the user's Claude account carry over. Tasks range from scheduling a meeting to dashboards, financial or sales forecasts, prototypes and web apps. Anthropic says it is built around Claude Opus 5.5, and TechCrunch names the same model. It can draw on the Claude Marketplace catalog of over 6,000 connectors and plugins; the post gives Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, DocuSign, HubSpot, Canva and Shopify as examples of what it can use.
The safety design is the most detailed part of the story. Before sending a message, making a payment, signing a document or deleting anything, Duet shows the action and waits for approval. Trading securities and moving money between accounts are not possible even with approval. Purchases go through Stripe's Link wallet, which uses a virtual card that never shares the real card number, or through Shop Pay at online stores, with approval. Credit card numbers and government IDs are blocked, and health or personal finance information is shared only if the user explicitly asks. Duet never directly sees passwords or sign-in tokens, and its actions are approved by Anthropic's auto mode classifier. Anthropic lists red-teaming, a penetration test by an outside firm and a public bug bounty program. Saved 1Password logins and a richer visual store editor for Shopify are listed as coming soon.
Agents move from chat to a small-business back office
Duet's framing shows agents shifting from a tool that answers questions to a background worker that briefs you, monitors your tools and can pay with approval. The audience is explicitly small business owners and professionals. Amin Ebrahimi of Shahs of Kabob, a small business tester quoted in Anthropic's announcement, describes the experience by saying "It brings it all together". TechCrunch names several competing assistant apps, among them OpenAI Dots, Instinct, Hark, Fo and Grok Bot.
The limits are the instructive part. The approval gates and the list of forbidden actions are a design a business can copy when it builds its own agent workflows. Anthropic also lists its own limitations: Duet can misread a thread or misjudge urgency, handles personal errands such as booking tickets less well than work tasks, cannot yet show its browser, and lives in a separate app from the main Claude app. Access comes through standalone iOS and Android apps and claude.ai, by waitlist, with US Pro and Max subscribers first; Team and Enterprise plans are not part of the preview yet. Duet comes with those two plans and draws on the same plan limits as the rest of Claude: its background checks count toward them, and larger tasks need more model turns and background work.
The sources say nothing about Türkiye, so this section is commentary
Neither Anthropic's announcement nor its product page mentions Türkiye, and TechCrunch's article has no reference to it either. Anthropic says the rollout begins with US Pro and Max users and names no other country. This section is therefore not a fact drawn from the sources but explicitly UNALSOFT commentary.
Our reading: for a business in Türkiye there is no concrete access today, and availability there is not addressed in the sources. What we could not learn from them: any date for Türkiye, the size of the waitlist and numeric values for the usage limits. The approval-gate and forbidden-action logic is still a design idea that can be judged apart from the product. We covered a related agent story in our Walmart agent gateway article.
The UNALSOFT view
We read this less as a product launch and more as a design pattern: handing work to an agent is possible, but every step that cannot be undone sits behind human approval, and some actions are off the table entirely. If you plan to use agents in your own workflows, ask the same questions: which actions need approval, which should never happen, and which tools the agent connects to with what authority. In our agentic AI work we write those boundaries down before the workflow is built. Because the sources carry no data about Türkiye, this article proposes no local access or legal conclusion.
Which tasks will you hand to agents, and with what approval?
A short conversation is enough to map the approval points and the forbidden actions in your workflow.