Anthropic Introduces Its Own Personal Agent: Claude Duet Follows Up on Work Proactively and Keeps Going Even After You Close Your Laptop

Anthropic has introduced Claude Duet, a standalone Agent application that proactively follows up on work and executes tasks on your behalf.

Key features include:

• Proactive follow-ups and morning briefings: Automatically reviews key updates across connected apps every morning to generate a summary; periodically checks connected apps and sends notifications only when needed.
• Continuous operation away from your device: Tasks run in Anthropic-hosted standalone cloud virtual machines and browsers, so shutting down your phone or closing your laptop will not interrupt ongoing work.
• Human confirmation for high-risk actions: Autonomously inspects cross-software information, analyzes data, and prepares drafts, but pauses to wait for your approval before sending external emails, settling merchant payments, signing documents, or deleting data.
• Inherits existing Claude memory and skills: Directly carries over memory, skills, and connected apps from your existing Claude account, unifying quick requests and long-horizon tasks within a single ongoing conversation (supporting both typing and voice).

It spots what needs attention before you do

When using standard conversational AI, humans usually notice an issue first and initiate the dialogue. Claude Duet proactively approaches you with items that require handling and continues getting the job done while you step away.

Duet is powered by Claude Opus 5.5, determining which items can wait, which can be prepared in advance, and which require your decision.

According to early tester feedback released by Anthropic, this proactivity is already reflected across several business scenarios:

• Advance warnings for calendar oversights: Real estate broker and agency owner Jess receives a daily briefing prepared by Duet each morning. One day, during a routine calendar check, Duet caught a property showing scheduled on the wrong date and warned her in time. Jess noted this was exactly the kind of thing she wanted an AI to catch proactively without her having to ask.
• Synthesizing information across multiple business systems: Amin, who manages restaurants with about 110 employees across South Florida, uses Duet for accounting, customer support, and marketing while reconciling invoices store by store. He noted that Duet pulled these responsibilities together, so "I don't have to open 100 tabs anymore."
• Advance suggestions for schedule conflicts: Brand strategist Mya had a scheduled call with a prospective client. Duet prepared a background briefing for the call and noticed the meeting bumped up against her flight, proactively suggesting that she reschedule.

Mya received a rescheduling recommendation; whether to adjust the calendar remained her decision.

Why work progresses after going offline or powering down

The key lies in the fact that its computing environment does not run locally on your phone or laptop. Once you assign tasks in Duet, execution is handed over to cloud virtual machines equipped with standalone cloud browsers:

• Opus 5.5 determines the steps: It interprets business goals, plans execution steps, and decides when to call tools and when to pause and ask for input.
• Cloud tools handle concrete execution: Inside cloud virtual machines and standalone browsers, it fetches application data, carries out specific actions, and prepares documents, charts, dashboards, financial forecasts, or web applications.

Phones and laptops merely serve as windows for assigning tasks and receiving key notifications. Even if your phone is turned off, your laptop closed, or you temporarily lose connection, the cloud virtual machine keeps advancing the task in the background; once your device reconnects, you receive notifications when the task is done or requires your authorization.

Proactive checks are not unconstrained real-time monitoring

Duet's proactivity does not mean continuous real-time background monitoring.

Anthropic says Duet checks connected apps every few hours. This interval is used primarily because continuous checking would deplete subscription plan usage limits much faster. As for when it messages you, model judgment determines what requires your attention and what can wait—a calibration Anthropic continues to refine based on testing feedback.

This also means that if an unexpected change or urgent message falls between two check intervals, Duet might miss it. You can tell Duet directly what you care about most and how frequently you want to receive notifications.

Permission boundaries: what it handles autonomously vs. what it never touches

When connecting an AI to core business tools, the biggest concern is often whether it might take unauthorized actions or leak data. Claude Duet establishes distinct permission boundaries across different actions and data.

1. Action permission tiers

• Autonomous execution: Reads information, compares and analyzes data, and drafts messages or prepares plans on its own.
• Requires human approval: Before sending outbound messages, making payments, signing documents, or deleting any content, Duet displays the specific upcoming action and waits for your approval. For example:
  • Sending emails or Slack messages;
  • Signing documents or contracts;
  • Deleting content;
  • Merchant purchases: When using Stripe's Link wallet, Duet first confirms the items with you, after which Link prompts for payment authorization and completes the transaction using a virtual card to avoid exposing actual card numbers; online checkout through Shopify's Shop Pay is also supported.
• Strictly prohibited actions: Certain actions will never be executed by Duet even if authorized by the user, including transferring funds between accounts and securities trading. Paying a merchant for a purchase is a different operation from transferring funds directly between accounts.

2. Sensitive information safeguards

As an Agent reviews materials and drafts outbound communications on your behalf, it must continually decide where information flows. Duet incorporates a dedicated new classifier and safety system prompt to check every message and guard against sensitive data leaks:

• Highest sensitivity data: Designed to reject sending information such as credit card numbers and government-issued identification outright.
• Other sensitive data: Information such as healthcare records or personal financial status is permitted only when the user explicitly requests sending it.

3. Credential isolation and security defenses

For credential storage when connecting enterprise apps, Duet follows the same design principles as Claude Managed Agents: the model itself does not directly see your login passwords or sign-in tokens. Under Managed Agents, passwords and credentials are held outside the execution environment; when the Agent calls tools, intermediate services handle authentication without ever exposing passwords to the model.

In addition, Duet incorporates targeted defense mechanisms:

• Prompt injection defense: Integrates defensive research from browser-use and related domains to guard against hidden instructions embedded in external web pages or emails attempting to derail the Agent or exceed authorized scope;
• Action classifier (Auto mode classifier): Prevents Duet from carrying out unauthorized steps that the user did not instruct.

Anthropic conducted red-teaming, hired an external security firm for penetration testing, and enrolled Duet in its public bug bounty program. While these measures serve to reduce unauthorized access and leakage risks, users must still review critical outputs in actual use.

The real boundaries of app connectors

Claude Duet can access over 6,000 connectors and plugins available in the Claude Marketplace.

Connectors allow the Agent to interact with corresponding external applications once authorized. The "6,000+" figure counts connectors and plugins available in the marketplace; it should not be taken to imply that every operation across every application can be completed smoothly.

Across specific business scenarios, Duet can support several typical workflows:

• Outstanding invoices and payment reminders: Retrieve unpaid invoices in Intuit QuickBooks or Stripe and draft payment reminder notices, sending them out only after your review and approval;
• Payroll and HR calculations: Log hours, bonuses, and expense reimbursements in Gusto, preparing payroll runs for human review and disbursement;
• Client follow-ups and contracting: Draft follow-up emails using client and deal histories from HubSpot; draft job descriptions on LinkedIn and publish upon approval, or prepare agreements in DocuSign for signature;
• Marketing design and e-commerce storefronts: Turn new product launches into social promotional assets using Canva; build and manage storefronts in Shopify. Official documentation notes that an enhanced visual editor for Shopify is coming soon, allowing page previews before going live.

Additionally, official documentation confirms that support for logging into websites using credentials saved in 1Password—with user permission—is coming soon. Users can also revoke authorization for any individual app at any time.

Eligibility, platform availability, and usage limits

For individual professionals and small-business owners planning to try it out, the current access terms are as follows:

1. Eligibility and platform support

• Current stage: In Research Preview; access requires signing up via the official waitlist.
• Supported plans: Initially available exclusively to Claude Pro and Claude Max individual subscribers located in the United States. Team and Enterprise plans are not currently included in this preview.
• Supported platforms: Dedicated native mobile applications for iOS and Android; accessible on desktop computers via the claude.ai web interface.

2. Usage limits and consumption

During the research preview, Claude Duet is included with Pro and Max subscription plans and shares usage limits with other Claude features. Its consumption profile differs from standard chat:

• Background proactive checks consume usage: Routine periodic checks of connected apps running in the background count against your plan's usage limits.
• Complex, large tasks involve multi-step work: When handling substantial tasks like building a financial forecasting model or assembling a presentation, multiple rounds of model interactions and tool operations run in the background, consuming larger portions of usage limits.

Notification preferences control what updates you receive and how often; the background checks themselves still count toward usage.

Start with tasks that have easy-to-verify outcomes

As an early research preview, Claude Duet still exhibits several practical limitations:

• Potential misinterpretations and priority misalignments: It may misread lengthy conversation context, miscalibrate urgency, or miss updates that occur between proactive check intervals.
• Core capabilities built around workplace scenarios: Designed primarily for professional work, with comparatively limited capability for personal errands like booking tickets.
• No direct view of the built-in browser: Duet cannot currently show you the live screen of its cloud browser, though users can inspect executed task steps and generated deliverables.

In practice, whether Duet truly saves effort comes down to how frequently omissions and misjudgments occur, along with the time and effort needed to manually review and approve actions. During early use, it is best approached by starting with well-defined tasks whose results are straightforward to verify.