A missed customer email, an overdue invoice and a meeting scheduled at the wrong time can each create work that nobody planned to do. Anthropic wants its latest app to spot those problems and start handling them before a business owner or employee thinks to ask.
The company on Wednesday introduced Claude Duet, an AI assistant that monitors connected applications, surfaces priorities and carries out assignments in the background. Anthropic designed the product around its Opus 5.5 model and is opening a research preview to U.S. subscribers on its Pro and Max plans through a waitlist.
Duet operates through a separate mobile app for iOS and Android, as well as claude.ai. It can prepare a morning briefing, investigate unpaid invoices, draft customer follow-ups, build financial forecasts and create web applications, according to Anthropic. Its remote computing environment lets it continue working after someone closes their laptop or turns off their phone.
For enterprise technology leaders, the announcement offers a concrete look at how Anthropic wants AI agents to participate in everyday business operations: drawing information from multiple systems, deciding what deserves attention and preparing actions for a person to approve.
But the initial release targets professionals and smaller businesses. Anthropic’s announcement does not extend Duet access to Team or Enterprise subscriptions, announce a Duet API or specify an enterprise rollout schedule.
An assistant that initiates work
Duet’s central proposition is that users should not have to initiate every interaction.
Anthropic says the app produces a daily briefing from connected services, then checks those accounts periodically for developments that warrant a notification. Users can adjust which subjects it flags and how frequently they want updates.
Those checks happen every few hours. Anthropic says more frequent monitoring would consume subscribers’ usage allowances faster, an important qualification for businesses considering time-sensitive workflows. Duet’s background activity does not amount to continuous, real-time supervision.
Users communicate through text or voice in a persistent conversation. Duet brings over account memory, skills and existing app connections, allowing it to begin with context the user has already supplied to Claude.
Underneath that interface, Duet runs in a virtual machine equipped with a browser. It determines which applications and steps a requested assignment requires, performs the work remotely and notifies the user when it finishes or needs a decision.
That architecture could make it useful for work that spans several systems: gathering account information, producing a document and preparing a follow-up, for example. The practical test will be whether it reliably preserves the relevant details as it moves between them.
How Duet differs from Remote Control and existing Claude apps
Claude users already have several ways to work from a phone, and those experiences overlap with Duet.
Claude Code Remote Control connects the mobile app or a browser to a coding session executing on the user’s computer. The phone provides access to that session; the host machine still supplies code execution and local file access. Remote Control supports Pro, Max, Team and Enterprise subscriptions, with organizational enablement required for the latter two.
Dispatch, an earlier Cowork feature, lets users send assignments from a phone to Claude on their desktop. It requires that computer to remain awake with the desktop app running. Anthropic now says Dispatch is closed to new users, although existing users can continue using it.
But cloud execution alone does not distinguish Duet from the current Claude app. Anthropic’s updated Cowork documentation says new Pro and Max tasks have run on its servers since October 6. Those tasks continue with the laptop shut, and scheduled jobs require no device to remain online. Anthropic is also combining chat and Cowork into one Claude experience.
Duet’s distinguishing proposition is the combination of a separate app, a continuing conversation, routine checks across connected accounts and judgment about when to interrupt the user. Remote Control provides access to a local coding agent; cloud Cowork executes requested or scheduled assignments; Duet aims to notice work that deserves attention in the first place. That is a product emphasis, not evidence that every underlying capability is exclusive to Duet.
From overdue invoices to payroll and purchases
Anthropic says Duet can draw on more than 6,000 connectors and plugins available through the Claude Marketplace.
Its announced workflows cover several business functions. Duet can examine QuickBooks or Stripe records for outstanding invoices and prepare reminders, use HubSpot’s customer information to draft outreach, and help assemble agreements for review and signature through Docusign.
For people operations, Anthropic describes preparing and running Gusto payroll with employee hours and adjustments, including bonuses and reimbursements. Duet can also draft recruitment listings and publish them to LinkedIn after a user reviews and authorizes the content.
Marketing and commerce examples include developing Canva campaigns and helping configure or maintain a Shopify store. Anthropic says a more capable storefront design interface, including a preview before publication, will arrive later.
The product also supports purchases through Stripe’s Link wallet and Shopify’s Shop Pay. In the Link flow, Duet checks the proposed item with the user, Link requests spending authorization, and payment uses a virtual card that keeps the underlying card number private.
Anthropic separately plans to let Duet use credentials stored in 1Password with user permission. That capability is forthcoming.
For technology buyers, the breadth of these integrations creates an obvious opportunity to reduce the manual handoffs between business applications. It also makes the permissions attached to each connection consequential: access sufficient to summarize an invoice is different from authority to initiate a payment.
Human approval remains part of the workflow
Anthropic draws an explicit boundary between preparing work and taking consequential action.
Duet can inspect information and assemble drafts independently. Before transmitting a message, paying for something, signing an agreement or deleting material, it presents the intended action and requests approval, according to the announcement.
Some activities remain prohibited even with authorization, including securities transactions and transfers between financial accounts. Users can also revoke an application connection.
Those distinctions matter when evaluating the payroll and purchasing examples. Anthropic is proposing software that can prepare and execute business workflows within approval constraints, rather than granting it unrestricted control over company spending.
The company also describes a dedicated sensitive-information classifier, paired with safety instructions, that evaluates messages to help prevent inappropriate disclosure. Anthropic says the system blocks sharing of its most restricted information categories, including payment-card numbers and government-issued identifiers. It permits disclosure of certain other sensitive material, such as health and personal financial information, only after an explicit instruction.
Additional safeguards include credential isolation intended to keep passwords and authentication tokens out of the model’s direct view, an action-approval classifier, and defenses against prompt injection — malicious instructions hidden inside content an agent encounters.
Anthropic says internal security teams and an outside firm tested Duet’s infrastructure, and its public bug bounty covers the product. These are descriptions of safeguards, however, rather than evidence that the agent will avoid every unauthorized action or information leak.
Pricing includes a background-work trade-off
Anthropic includes Duet in Pro and Max subscriptions during the preview, with no separate Duet charge announced. Its published subscription pricing lists Pro at $20 monthly, or $200 for an annual subscription, and Max starting at $100 monthly.
Subscribers share their existing Claude usage allowance with Duet. Automated checks consume part of that allowance, while assignments such as constructing a financial model or presentation require additional work.
That creates a budgeting question beyond the subscription price: how much useful output can a user obtain after accounting for the agent’s recurring monitoring?
Anthropic does not provide task-level cost estimates, quantified productivity gains or benchmark results in the Duet announcement. Organizations assessing the preview will therefore need to measure completed work, correction time and allowance consumption themselves.
The initial plan eligibility also matters for governance. Anthropic’s pricing page lists organization-level features such as audit logs and custom retention controls under Enterprise. Buyers should not assume that the Duet preview on an individual subscription carries the same controls.
Where Muse, Dots, Instinct and Tolan fit
Duet enters an existing market for persistent personal agents. Its closest comparisons include Meta’s Muse and OpenAI’s Dots, both of which already promise proactive assistance and execution on a dedicated cloud computer.
Meta introduced Muse in September with an emphasis on everyday tasks, purchases, travel and longer-term goals. Users reach it through its app or WhatsApp. Like Duet, it continues working after the app closes and seeks permission for sensitive actions. Meta also describes an activity history and a separate monitoring agent that checks outbound actions. Human approval and cloud isolation therefore do not, by themselves, differentiate Duet.
Muse’s rapid uptake helps explain the competitive stakes: Sensor Tower estimates, reported by Yahoo Finance on October 20, put the app at 10.9 million downloads since its September 8 debut, including 10.2 million in the U.S., and 2.5 million daily active users on October 16. That early traction suggests consumer appetite for AI that goes beyond answering questions to carrying out tasks—a shift now taking clearer shape with OpenAI’s Dots and Anthropic’s Duet.
OpenAI’s Dots overlap particularly closely with Duet’s professional ambitions. OpenAI describes agents powered by GPT-6 Astra that investigate issues, build applications and progress projects between conversations. They work through ChatGPT and messaging channels including Slack and Teams. OpenAI’s access documentation includes Pro and Business Premium subscribers, plus an administrator-enabled Enterprise beta. That gives Dots a stated organizational access route that Duet’s initial announcement does not offer.
Instinct competes on delegation through familiar communication: users text or call an assistant that can operate a phone and computer to handle errands. Its examples emphasize groceries, transport and bookings. It comes from Spear Street Technology, a separate company, rather than Meta.
Tolan, made by Portola, is a more indirect comparison. Its character-based experience emphasizes conversation, memory, check-ins and help with everyday goals. Those qualities overlap with a personal assistant, but its public materials do not establish an equivalent system for operating payroll, customer-management tools or other business applications. It should not be treated as a demonstrated substitute for Duet’s business workflows.
The pricing also differs substantially:
Duet’s $20 monthly entry point undercuts the $100 minimum individual Pro subscription for Dots, but Muse offers a free entry point. Existing subscribers face a different calculation because Duet and a first Dot come bundled with broader subscriptions.
For enterprise buyers, the meaningful comparison extends beyond whether an agent can answer from a phone. It includes whether it can access approved systems, explain proposed changes, preserve a reviewable history and finish useful work within a predictable allowance. Duet’s launch makes those comparisons possible, but provides no head-to-head evidence that it completes the same assignments more accurately or cheaply.
Early examples, with reliability still to prove
Anthropic’s launch examples emphasize small operational improvements that could accumulate throughout a workday.
The company describes real estate broker Jess Provencher using Duet to catch a property showing entered on the wrong date. Restaurant owner Amin Ebrahimi, whose Shahs of Kabob business has approximately 110 employees, has used it for accounting, customer service and marketing, according to Anthropic.
“It brings it all together,” Ebrahimi said in the company’s announcement. “I don’t need to open up 100 tabs.”
These are company-selected customer accounts, not independent evaluations of accuracy or business impact. Anthropic acknowledges that the preview can misunderstand conversations and assign the wrong urgency to an issue. It also says Duet cannot yet display its browser to the user, limiting visibility into that part of its work.
For enterprise teams, a sensible initial evaluation would focus on bounded tasks whose results staff can readily check: briefing preparation, invoice follow-up drafts or meeting coordination. More consequential workflows would require closer examination of approval behavior, permissions and how staff can reconstruct what happened when something goes wrong.
Duet gives businesses a way to test whether an assistant that initiates work can reduce the attention required to manage it. Its value will depend on how often those interventions help — and how much supervision remains necessary to trust the results.
Want to get more out of your agents with fewer tokens and less context? Apply to attend VentureBeat’s Agentic Infrastructure Stack Summit, October 28 in San Francisco.