Claude Duet Joins The Race To Keep AI Agents Always On

🔍 Read the full analysis: Claude Duet Joins The Race To Keep AI Agents Always On on ThorstenMeyerAI.com

Anthropic launched Claude Duet, a waitlisted research preview of an AI agent that can work on a cloud computer and send users updates when a task needs attention. The service is available to U.S. Claude Pro and Max subscribers by waitlist, and matters because it extends Claude beyond a chat window into connected apps and background work.

Anthropic describes Duet as a separate app available on iOS, Android and claude.ai. It says the agent can prepare a morning brief, accept tasks by text or voice, and work in an ongoing thread while the user’s own computer is off. Examples include scheduling meetings, creating dashboards, building sales forecasts and making small web apps. Duet runs tasks on a virtual machine with its own browser and draws on the memory, skills and connected apps associated with the user’s Claude account.

The company says Duet can use more than 6,000 connectors and plugins in the Claude Marketplace, including services used for accounting, payroll, contracts and online stores. Anthropic’s examples include finding unpaid invoices in QuickBooks or Stripe, working with payroll in Gusto, handling agreements through DocuSign and assisting with Shopify storefront tasks. These are capabilities described by Anthropic; the source material does not report independent testing of the preview.

Anthropic’s October 23 safety post says Duet must ask before sending potentially sensitive messages, making payments or deleting items. It shows the proposed action, destination and content, including recipients, and waits for approval. Anthropic says an additional check runs separately from the working model to assess whether an action fits the user’s intent. The company also says trading securities and moving money between accounts are blocked, even if a user requests them. Purchases use Stripe’s Link wallet and a one-time card limited to a merchant and amount, according to Anthropic.

Background Agents Put Access in Focus

Duet adds Anthropic to a developing contest over assistants that can act while users are away, rather than responding only when prompted in a chat. The practical appeal is delegating routine work across connected services; the corresponding risk is that an agent may act on sensitive information or take an unwanted action. Anthropic’s approval gates and blocked-action list are therefore central to how the preview is presented, not peripheral product details.

The service also makes the connected account and the interface a point of competition. Users may care not only which model performs a task, but which assistant holds their app access, task history and approvals. The supplied source describes a broader market in which assistants may route work to models from other companies. That could make control over the agent and its connections matter alongside the model itself, though the available material does not establish how that market will develop or how much choice customers will have.

For businesses, the near-term question is operational: whether background automation can save time without creating unacceptable access, oversight or reliability problems. A limited pilot using read-mostly permissions can test those trade-offs before an organization connects high-impact systems. This is a practical recommendation from the source, not a claim that Duet has already shown measurable productivity gains.

A Crowded Race for Agents

Claude Duet follows other agent announcements listed in the source material: SpaceXAI opened a Grok Bot beta on August 11, Meta launched Muse on September 8, OpenAI introduced Dots on September 29, and Google released Gemini Agent on October 8, according to TechTarget as cited there. Anthropic announced Duet on October 21. The products are described as sharing a broad premise: an assistant can use a cloud-based computer, retain context and return with results or requests for attention.

The source identifies Duet as built on Claude Opus 5.5, though the provided material does not independently verify the model or compare performance among the products. It also reports that Grok Bot may route tasks to other providers‘ models, including Claude Opus 5.5, citing a post by Elon Musk and an xAI security FAQ. In the FAQ wording supplied, Grok Bot’s model choice cannot be restricted through the team model allowlist, and there is no customer-facing model picker. This is relevant market context, but does not establish that Duet itself routes work to competing models.

Preview Limits and Open Questions

Duet is a research preview, and the supplied material says no independent test has yet established whether its controls work as described in real use. Anthropic says the early version may misread message threads or misjudge urgency, is weaker at personal errands such as booking tickets, and cannot yet show users its browser. It checks connected accounts every few hours rather than continuously, which means it may miss developments between checks.

Anthropic also says background checks count toward usage limits shared with the rest of Claude, while large jobs such as financial models require more model turns. Disconnecting an app stops future access, but information Duet has already learned remains until the user resets its memory. The source does not specify the length of the waitlist, a broader release date, the full set of supported regions, or whether the stated limits and controls will change during the preview.

Other details need verification as users try the product: how consistently approval prompts appear, how well the separate classifier handles ambiguous requests, and how Duet behaves when connected services return incomplete or conflicting information. The source reports Anthropic’s statements on credentials and sensitive data, including that Duet does not see passwords, browser cookies or sign-in tokens; it provides no independent audit findings.

Testing Duet Before Wider Access

Anthropic has opened access through a waitlist for U.S. Pro and Max subscribers; the supplied information does not give a date for wider availability. The next useful evidence will come from how the preview performs in ordinary tasks, how often it requests approval and whether users can understand and manage the information it retains.

Companies considering an evaluation can begin with a small, documented pilot, limit permissions to read-mostly access and define an exit plan, as the source recommends. Individual users can test low-stakes tasks, connect only a few apps and retain approval requirements for actions that send, pay or delete. Those steps reduce exposure but do not replace review of Anthropic’s current terms and controls.

Readers should watch for updates to Duet’s supported apps, account eligibility, usage limits and safety measures, as well as independent testing of its performance. Until such evidence is available, Anthropic’s launch details describe the product’s intended operation, not proof that every task will be completed reliably or safely.

Key Questions

What is Claude Duet?

Claude Duet is Anthropic’s research-preview agent, described as working on a cloud computer, using connected apps and contacting users when a task needs attention.

Who can use the preview?

According to the supplied source, access is by waitlist for U.S. Claude Pro and Max subscribers. No public release date is provided.

Can Duet send messages or make payments without approval?

Anthropic says Duet must show the proposed action and wait for approval before sending potentially sensitive messages, paying or deleting items. The company says trading securities and moving money between accounts are blocked.

What limitations has Anthropic disclosed?

Anthropic says Duet may misread threads or urgency, is weaker at some personal errands, cannot yet show its browser and checks connected accounts every few hours. Background work also uses the plan’s shared usage limits.

Has Claude Duet been independently tested?

The supplied source says no independent test exists yet. The current details about features and safeguards are Anthropic’s statements about a research preview.

Source: ThorstenMeyerAI.com