Claude Duet: Anthropic's proactive agent arrives as a research preview

On October 21, 2026, Anthropic announced a research preview of Claude Duet, an app where Claude acts as a proactive agent: throughout the day it monitors connected apps and only messages the user when a decision is needed. The official announcement, published on claude.com and shared by the @claudeai account on X, specifies that the app is separate from Claude's main interface and runs on iOS, Android and the web (claude.ai). The agent is built on Claude Opus 5.5 and uses its own virtual machine with a built-in browser, so the user's device doesn't need to stay on. (operational unknown: it's unclear how much background activity eats into usage limits)

Access is currently limited to a waitlist, starting with Pro and Max subscribers in the United States. No dates have been given for a rollout to the rest of the world or for general availability, so the app isn't available in Italy yet. Anthropic says Duet is included in the Pro and Max plans at no extra cost, but it draws on the same usage limits as the rest of Claude: background checks count too, and bigger tasks burn through the limits faster. (documented gap: there are no independent metrics on the real cost of background checks)

Duet promises to handle professional tasks such as scheduling meetings, building dashboards, preparing financial or sales forecasts, and prototyping and building web apps, using more than 6,000 connectors and plugins from the Claude Marketplace, including Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, Docusign, HubSpot, Canva and Shopify. Before sending a message, making a payment, signing a document or deleting anything, the agent shows the planned action and waits for the user's approval; securities trading and money transfers are off-limits even with consent. (operational unknown: plugin coverage is stated by Anthropic but hasn't been verified by third parties)

On security, Anthropic points to a new classifier and a system prompt designed to prevent leaks of sensitive information. According to Anthropic, the most protected tier (credit cards, ID documents) is meant never to be shared, while health and financial data are shared only on explicit instruction. The company also cites an "auto mode" classifier on actions, internal red-teaming, an external penetration test and a public bug bounty program. However, there is no independent data confirming that these mechanisms work. (operational gap: no public results from the security tests).

According to Digital Trends, the product targets professionals and small business owners and is Anthropic's answer to agents like OpenAI's "dots". In the same month, Meta and Sierra proposed a protocol for personal agents.