Claude Duet Launches: Anthropic Opens Preview of Proactive AI Agent
The launch lands the same day Meta reported Q3 results and Zuckerberg claimed more than 10 million Muse setups — and it is the third major launch of a proactive personal agent in roughly six weeks.
Anthropic opened a research preview of Claude Duet on Wednesday, October 21, 2026 — an "always on" agent that monitors connected apps, surfaces what needs attention, and carries out tasks in the background.
What Duet Does
According to Anthropic's blog post, reproduced by CNBC, the company describes the product this way: "It comes to you with what needs your attention, completes tasks you ask it to do, keeps working even when you step away, and checks with you before it acts on your behalf."
Duet is built on Anthropic's Opus 5.5 model and runs in a virtual machine equipped with a browser, which the company says allows the agent to keep working even when the user's devices are switched off. The agent checks connected accounts roughly every few hours; Anthropic says the checks happen periodically in part to conserve the usage quota. More frequent monitoring would burn through the allowances faster.
The integrations run through Claude Marketplace, which according to Anthropic's figures, as reported by Reuters/Miami Herald, counts more than 6,000 connectors, including Gmail and Slack. Duet can also make purchases via Stripe's digital wallet Link and Shopify's Shop Pay — details that come from the company itself and have not been independently verified.
Who Gets Access — and What It Costs
The preview is limited to users in the US on Anthropic's Pro and Max subscriptions, with access via a waitlist. MacRumors reports that Duet can be used via the web or a standalone iOS app, and that usage counts toward the existing Claude limits — there is no separate pricing scheme for Duet.
Individual details from Reuters/Miami Herald and VentureBeat's headline also suggest Android access, but this is not confirmed across sources and should for now be considered unresolved.
Safety Gates and the Company's Own Caveats
Anthropic says Duet requires user approval before sending messages, making payments, signing documents, or deleting files. The company further states that the agent does not see passwords or login details when it is given account access, and that a dedicated classifier and a safety prompt in the system instructions check every message and are meant to protect against leakage of sensitive information such as credit card numbers and ID details. All of this consists of company statements — none of the safety mechanisms have been independently tested.
Anthropic itself also tempers expectations. The company acknowledges that the agent "may misread a thread or miscalibrate what's urgent," and that since Duet is primarily designed for the workplace, personal errands like booking tickets are "not quite as strong."
Mike Krieger, co-lead of Anthropic's Labs team, said in a launch video, according to Business Insider, that the agent "can have the judgment and discretion needed to know when to be proactive, and when to act on your behalf."
Workplace First — in a Three-Way Race
The timing is no accident. That same Wednesday, Meta reported Q3 results, and Zuckerberg told analysts that more than 10 million people in the US and Canada "have set up Muse personal agents" over the past month and a half. That is a company statement from an earnings presentation, not an independent measurement. The data firm Apptopia estimates — as a third party — that Muse had more than 14 million installations as of Wednesday. Apptopia has, according to Business Insider, also found that Claude users who installed Muse used Claude more than before, but fewer minutes per day than on Meta's product; the methodology behind this has not been published.
With OpenAI's Dots, launched at the developer conference earlier this fall, and Meta's Muse in September, Duet is the third major proactive personal agent in roughly six weeks. But while Muse is consumer-oriented, Anthropic positions Duet workplace-first — and the bet is led by someone who knows both sides: Karan Ahuja, who led product management at Meta Superintelligence Labs until July and started at Anthropic in August, has led Duet's development. On X, he wrote that the team built it "for professionals and small business owners juggling a lot between work and home."
What Was Not Announced
VentureBeat points out what Anthropic's announcement leaves out: Duet is not getting access for Team or Enterprise subscriptions, no Duet API was announced, and there is no timeline for enterprise rollout or international availability. That is worth noting in light of Anthropic's broader agent landscape — Remote Control, the now-closed Dispatch, and cloud-based Cowork tasks that have been running on servers since October 6 — as Duet is a consumer- and pro-oriented product from a company that has otherwise prioritized enterprise customers.
Open Questions
Much remains uncertain: how the approval gates work in practice when the agent is supposed to "exercise judgment" about what is urgent; how the plan limits hold up against frequent background checks in everyday use; whether an Android version actually exists; and when — or whether — Duet reaches enterprise customers and markets outside the US. With three proactive agents launched in six weeks, the competitive pressure is clear, but which of them users actually entrust with watching their accounts remains an open question.