Anthropic’s Claude Duet bets on an agent that finds the work

A different kind of Claude agent

Duet sends a daily summary of connected services, then checks them periodically and alerts users to important changes. Users choose what it monitors and how often they receive notifications. Checks run every few hours, not continuously; more frequent monitoring would use up subscription limits faster. That makes Duet less suited to organizations that need immediate responses.

Users interact with it by text or voice in one ongoing chat. It can draw on saved account memory, skills and app connections, then use a virtual machine with a browser to carry out work remotely. For example, it might gather account information, prepare a document and draft a follow-up email.

That workflow overlaps with other Claude products, but the distinction is in how work starts and where it runs:

That last combination is Duet’s product emphasis, not proof that its capabilities are exclusive to Duet. The question is whether it can move reliably between services while retaining the details a task requires.

Broad access, bounded actions

Anthropic says Duet can use more than 6,000 connectors and plugins from Claude Marketplace. Its examples span finance, customer management, hiring, marketing and shopping:

Duet can research and prepare drafts, but it asks for confirmation before sending a message, making a payment, signing a contract or deleting material. Some actions, including securities transactions and transfers between financial accounts, are prohibited even with user consent. Users can revoke an app connection.

Anthropic also describes a sensitive-information classifier that works alongside its safety instructions. It checks messages and is intended to prevent inappropriate disclosure. The system blocks the most restricted categories, including payment-card numbers and government identifiers; some other sensitive information, such as health and personal-finance data, can be disclosed only after a direct instruction. Other safeguards include credential isolation, an action-confirmation classifier and protection against prompt injection.

Anthropic says internal security teams and an external company reviewed Duet’s infrastructure, and that its public vulnerability program covers the product. Those measures matter, but they do not establish that the agent will never take an unauthorized action or expose data. For a business, the relevant boundary is not just what Duet can connect to, but what each connection allows it to do.

The price of background work

During its research launch, Duet is included in Pro and Max subscriptions, with no separate price announced. Pro costs $20 a month or $200 annually; Max starts at $100 a month.

Duet shares its usage limit with other Claude features. Background checks consume part of that allowance, while work such as preparing a financial model or presentation requires additional computing. So the subscription price alone does not say how much useful work an organization will get. Anthropic has not published per-task costs, measured productivity gains or benchmark results for Duet. Teams evaluating the research version will need to track completed work, time spent correcting it and usage consumed.

Plan type matters too. Anthropic says Enterprise includes organization-level features such as audit logs and custom data-retention settings. Those controls should not be assumed to come with a preview of Duet on an individual subscription.

Duet also enters a growing market for agents that stay active and act without a fresh prompt. Meta’s Muse and OpenAI’s Dots are close comparisons: both promise proactive help and task execution on a separate cloud computer.

Meta introduced Muse in September for everyday tasks, shopping, travel and longer-term goals, through its app or WhatsApp. It keeps working when the app is closed and asks permission for sensitive actions. Meta also describes activity history and a separate agent that checks outgoing actions. Human confirmation and cloud isolation, then, are not distinctions unique to Duet.

Muse’s early adoption suggests interest in assistants that do more than answer questions. Sensor Tower data published by Yahoo Finance on October 20 put downloads at 10.9 million since its September 8 launch, including 10.2 million in the US. It had 2.5 million daily active users on October 16.

OpenAI positions Dots more directly around professional work. The company describes agents powered by GPT-6 Astra that investigate problems, build applications and move projects forward between conversations. They work through ChatGPT and messaging channels including Slack and Teams. OpenAI’s documentation lists Pro and Business Premium subscribers, plus an Enterprise beta that an administrator must enable. That gives Dots a stated path to organizational access that Duet’s initial offer does not yet have.

Instinct, from Spear Street Technology, focuses on requests through familiar communication channels: users can text or call an assistant that operates a phone and computer for everyday tasks such as grocery shopping, travel and bookings. Tolan, from Portola, is a more distant comparison. Its character-based interface emphasizes conversation, memory, check-ins and help with personal goals, but public materials do not show a comparable system for payroll, customer management and other business applications.

Pricing changes the comparison. Duet’s lowest entry point is $20 a month, below the $100 starting price for an individual Pro plan for Dots; Muse also has a free option. For existing subscribers, however, Duet and the first Dot are included in broader plans.

I think the more important comparison for businesses is not whether these agents can respond from a phone. It is whether they can work in approved systems, explain proposed changes, leave an auditable trail and deliver useful results within a predictable usage limit. Duet’s launch makes that comparison possible, but it does not show that the product performs the same work more accurately or cheaply than its competitors.

What the launch cannot yet show

Anthropic’s early examples are small improvements that might add up over a workday. The company says real estate agent Jess Provencher used Duet to catch a property showing scheduled for the wrong date. Restaurant owner Amin Ebrahimi, whose Shahs of Kabob has about 110 employees, used it for accounting, customer service and marketing. Ebrahimi said Duet brought tasks together in one place and saved him from opening a hundred tabs.

These are examples selected by Anthropic, not independent evidence of accuracy or business impact. The company acknowledges that the research version may misunderstand a conversation or misjudge urgency. It also says Duet cannot yet show users its browser, leaving some of the agent’s work out of view.

My guess is that early business use will make most sense for bounded tasks employees can check quickly: summaries, draft payment reminders or meeting coordination. Higher-stakes workflows will require closer scrutiny of permissions, confirmations and the ability to reconstruct what happened when something goes wrong. Duet’s test is not whether it can find work to do, but whether its interventions save more effort than the oversight they require.