Anthropic launches personal agent Claude Duet, opening waitlist applications to Pro and Max users in the US

On October 21 local time, Anthropic launched Claude Duet, a personal agent application, opening research preview waitlist applications to Claude Pro and Max subscribers in the United States.

The product is based on Claude Opus 5.5 and can connect to more than 6,000 apps and plugins. It proactively organizes to-do items, drafts replies, prepares payroll and financial materials, and continues executing tasks while the user is offline, requesting confirmation when it encounters sensitive operations such as sending emails or making payments.

The scenarios Anthropic demonstrated include distilling important items from emails and calendars every morning, identifying unpaid invoices and drafting payment reminder notices, as well as preparing financial models, presentation decks and business dashboards. Users do not need to stay in the chat window continuously; the agent contacts the user again when a task is completed or a decision is needed.

However, Anthropic also acknowledged that the research preview may still misjudge the urgency of information or miss new items in the gaps between background checks.

Before Claude Duet, products related to leading models such as MetaMuse, xAIGrok Bot and OpenAIDots had all begun to stake out the personal agent track, attempting to turn model capabilities into products that continuously execute tasks.

Leading domestic vendors are also accelerating. Alibaba's Qwen Go is embedded in the Qwen app and has opened experience applications, covering tasks such as claims processing, appointments, price comparison and food ordering; Tencent Yuanbao is developing a standalone app, "Yuanbao Assistant," to carry long-cycle tasks and long-term memory; ByteDance is reportedly set to launch a standalone personal assistant, "Xiaodou" (the Spell project), planned to go online as a standalone app and currently still in internal testing.

Gartner data shows that global spending on AI agents and assistants will grow from $16.481 billion in 2025 to $65.472 billion in 2027, roughly four times the original figure.

Behind the changes in product form is the model vendors' contest for the next-generation user entry point. Traditional chat assistants usually wait for users to ask questions, while task-oriented agents attempt to continuously understand user goals, retrieve cross-application information, formulate plans and execute operations.

Once users connect email, calendars, files and business systems to the agent, the product's value no longer depends solely on the quality of a single answer, but also on whether it can remember user preferences over the long term, understand work context and reliably complete multi-step tasks. For model vendors, this may bring higher-frequency usage, deeper application connections, and subscription and service opportunities for individuals and enterprises.

But as a result, the competitive moat for personal agents has also expanded from model capabilities to memory, tools and security architecture. Meta designed a user-specific secure cloud virtual machine for Muse, isolating the agent's runtime environment from credential storage, with an independent Sentinel component reviewing external operations; sensitive actions such as sending emails and shopping payments require user authorization, and during payment a one-time card number can also be generated via Stripe Link to reduce exposure of real payment information.

The more autonomously a personal agent can act, the more it needs access to users' private data and real accounts. Long-term memory, permission control, operation auditing, independent execution environments and confirmation of key steps will together with model capabilities constitute competitiveness.