Anthropic's new agent Claude Duet can pay and sign for you, once you approve

Every morning, a summary of what matters in your emails, messages and management tools. Then a message only when something needs your attention. That is the promise of Claude Duet, the AI agent Anthropic unveiled on October 21. European readers will have to wait: the preview is limited to the United States.

Duet is a standalone app for iPhone, Android and the claude.ai website. It is a "research preview" available through a waitlist, first to US subscribers on the Pro and Max plans. There is no extra charge, but background checks and long-running tasks count against the subscription's usage quota. The agent runs on the Claude Opus 5.5 model and has its own virtual machine with its own browser, so your computer does not need to stay on.

Pay, sign, delete: always after your go-ahead

Anthropic claims access to more than 6,000 connectors, including Gmail, Slack, QuickBooks, Stripe, Docusign and Shopify. Duet can schedule meetings, build dashboards or financial forecasts, and make purchases with a virtual card. Before sending a message, paying, signing a document or deleting anything, it must show exactly what it is about to do and wait for your approval, according to the company.

Some actions remain off-limits even with your approval: buying or selling securities, and transferring money between accounts. A filter is also supposed to block the sharing of credit card numbers and identity documents. Anthropic says the agent never directly sees your passwords or login tokens. It also says it has built defenses against prompt injection, the hidden instructions planted in a web page or an email to hijack an AI. An outside firm carried out penetration tests, and the company promises more security details "in the coming days." No independent analysis has been published so far.

What the company and an early tester admit

Anthropic acknowledges that Duet can still misread a conversation or misjudge how urgent a topic is. It also concedes that the agent is weaker at personal errands, such as booking tickets, and that it cannot yet show the user what its browser is doing.

Dan Shipper, of the US outlet Every, used it for two weeks with early access provided by Anthropic. His verdict is broadly positive: he sees it as the first personal agent he would call a "professional tool." He does note some limits. Scheduled tasks cannot use custom connectors. He finds its email drafts less convincing than those from OpenAI's competing agent. And continuous monitoring eats into the subscription quota.

Before handing your accounts to an agent

Duet arrives to compete with agents from Meta and OpenAI, as several major sites already restrict these assistants. JPMorgan, for instance, bans them from acting on its customers' accounts. For an SME tempted by this kind of tool, a few questions are in order:

• which applications can the agent read, and which can it modify?
• are the approvals it requests actually read, or accepted out of habit?
• do your vendors' terms of service allow a third-party agent?
• where is the data the agent remembers stored?

Anthropic has given no date for Europe. The next milestone will be the publication of the promised security details and the first independent tests.