Anthropic Enters the AI Agent War with Claude Duet

Anthropic announced Claude Duet on Wednesday, October 21, a new artificial intelligence agent designed for self-employed professionals and small business owners. The tool positions itself as a direct response to Meta's Muse and OpenAI's Dots, expanding a race that is reshaping expectations for digital assistance.

Unlike conventional versions of Claude — the chat that runs on claude.ai or the traditional mobile app — Duet operates proactively. Rather than waiting for users to type a question, the agent monitors connected applications, identifies priorities, and executes tasks in the background, notifying the user only when something requires attention or approval.

How It Works

Duet runs on a remote virtual machine with its own browser, meaning it continues operating even when the user's device is turned off. Each morning, it sends a summary of the most important items spread across connected services — email, calendar, CRM, and financial tools — and allows the user to adjust which subjects deserve alerts and how frequently.

The architecture behind this is the same one Anthropic has already been testing in other areas: "Pro" and "Max" tasks that have already been running on servers since October 6, with scheduled jobs and remote execution. Duet brings this same infrastructure but adds the ability to initiate interactions without the user needing to open the app first.

The announced integrations include over 6,000 connectors available in the Claude Marketplace — Gmail, Slack, QuickBooks, Stripe, HubSpot, Gusto, Docusign, Canva, and Shopify. The agent can analyze accounts receivable in QuickBooks, prepare payment reminders in Gmail, draft follow-up emails, build campaigns in Canva, and even set up stores on Shopify.

In the payment flow, Duet uses Stripe Link and Shop Pay. Before completing a purchase, the agent checks the proposed item with the user, Link requests spending authorization, and payment is made with a virtual card that protects the user's real card number.

The Architecture Behind Opus 5.5

Duet is built on Claude Opus 5.5, Anthropic's most capable model. This choice is deliberate: proactive agents require judgment about what is urgent versus what can wait, about when to notify the user versus when to simply act. Smaller models can complete isolated tasks, but the ability to prioritize and contextualize actions in real time is compromised.

What differentiates Duet from older automation approaches is the concept that the assistant "takes the lead" — not just responding, but anticipating. In a typical scenario, Duet identifies an overdue invoice in QuickBooks, consults the customer's purchase history in HubSpot, drafts a personalized message in Gmail, prepares a reminder, and delivers everything in a single notification with the option to send or edit first.

Security and Control

Anthropic emphasizes that security is a design pillar. Passwords and authentication tokens are never directly exposed to the model — the agent interacts with credentials in an isolated manner, following the same pattern adopted by Anthropic's Managed Agents. A new action classifier and a security prompt system verify each message to prevent the leakage of sensitive information, such as credit card numbers and government documents.

Anthropic's internal security teams and an external penetration testing firm tested Duet, which is also covered by the company's public bug bounty program. Defenses against prompt injection — malicious instructions hidden within content the agent encounters — were implemented based on the company's own research on browser use.

The Personal Agent Market

The Duet launch arrives at a moment when the three leading AI laboratories are all launching their own personal agents. Meta introduced Muse in September, focused on everyday tasks, shopping, and travel — the product had already received millions of downloads. OpenAI launched "Dots" the previous month, always-on agents operating within Slack and Microsoft Teams. Anthropic is entering slightly later, but with a different strategy: instead of competing directly with consumers on personal tasks, Duet targets professionals and small businesses.

This reflects a clear positioning decision. Meta's and OpenAI's agents already count with millions of users and have a more consumer-focused approach. Duet, for now, will not be available for Anthropic's Team or Enterprise subscribers, and there is no schedule for an API or enterprise launch. The initial objective appears to be consolidating a more productivity-focused user base before scaling.

What to Expect

Duet represents a turning point in how Anthropic envisions its agents being used — not as reactive tools, but as assistants that operate in the background. The question is not whether proactive agents will become standard, but which model will set the rules for how they behave in the real world.

With Duet, Anthropic is betting that the model's judgment — what to prioritize, when to notify, when to act — is the differentiator that separates a useful assistant from constant noise.

Sources: The Verge, VentureBeat, Business Insider

✓ Independent sources cross-checked and verified before publishing