Anthropic has opened a research preview of Claude Duet, a personal agent it describes as proactive rather than reactive: it comes to you with what needs your attention, completes tasks you hand it, keeps working while you step away, and checks with you before it acts on your behalf. The announcement, made Tuesday, frames Duet around professionals and small business owners “with more on their plates than they can track” — and it lands in the middle of the week America’s banks started blocking agents just like it.
What it does
Duet is built on Claude Opus 5.5 and ships as a standalone app on iOS and Android plus a web surface on claude.ai, included in Pro and Max plans with the same usage limits as the rest of Claude — access for now via a waitlist. The product page positions it plainly: “Get back to what needs you.”
The daily rhythm, per Anthropic: each morning Duet sends a short brief on the most important items across your connected apps; through the day it keeps tabs on those apps and messages you only when something needs your attention. Task-wise it spans everything from scheduling meetings to building dashboards with dozens of tools, financial or sales forecasts, prototyping and web apps. It works in the background through its own virtual machine and comes back when it needs approval.
The limits are as telling as the features. As TechCrunch reports, Anthropic is distinguishing Duet from agents like Instinct by admitting what it is bad at: errands and booking tickets are off the script, and it cannot show you its browser use as it happens. It connects to more than 6,000 connectors and plugins in the Claude Marketplace, and it can make purchases on a user’s behalf through Stripe Link or Shopify’s Shop Pay — but sensitive steps require approval, and it will not trade securities or move money between accounts.
How Anthropic is pitching the safety story
The announcement leans on infrastructure built for earlier agents: credentials handled the way Claude Managed Agents does it, so the agent never directly sees your passwords or sign-in tokens; actions gated by the same auto mode classifier that protects Claude Code; prompt-injection defences drawn from Anthropic’s browser-use research. New in Duet is a dedicated classifier and safety system prompt that checks each message — requests to share credit card numbers or government identification are rejected by default, and health or personal-finance information passes only when explicitly instructed. Anthropic says red teams and an external firm penetration-tested the product, with a public bug bounty on top.
A crowded race, entered deliberately
The timing is not accidental. CNBC notes Duet is Anthropic’s take on a field that already has runners: Meta’s Muse — which Meta says has more than 10 million users — OpenAI’s Dots, Hark, Instinct and xAI’s Grok Bot, as Engadget frames it. Anthropic is entering last among the majors, with the most conservative product brief.
That conservatism reads differently after this week than it would have a month ago. The agent ecosystem’s security debt is real — researchers tallied 28 million leaked secrets and 47,000 backdoored machines across agent infrastructure, and agents running real businesses on payment rails changed the industry’s questions in March. Amazon had already blocked Meta’s Muse shopping agent from its site in September. Duet’s design choices — approval gates on payments, no money movement, an agent that can’t see your passwords — are effectively Anthropic pre-conceding the position banks took this week: agents may look, but a human moves the money. The pitch “waits for your ok when it matters” is the productised version of the read-only truce.
The risk is that the market rewards the opposite. Muse got to 10 million users by moving fast on personal tasks, shopping included; Duet deliberately can’t book your tickets. If consumers vote for capability and enterprises vote for control, Anthropic is betting the professional market — the one that pays for Pro and Max — votes for control. Whether an agent that asks permission often enough to be safe, but not so often it becomes a nag, actually exists is the product question of 2027.
The New Zealand read
For now, Duet is a US research preview: Pro and Max subscribers in the US join the waitlist, and Anthropic has not announced availability elsewhere. New Zealand subscribers will wait, as they did for earlier staged rollouts. What arrives here eventually is the same design problem New Zealand’s own regulators have been writing about — the Reserve Bank and FMA told financial firms this month to assume AI incidents will happen, and a proactive agent that reads your email is exactly the kind of system those expectations were drafted for. The interesting detail for NZ observers is that Anthropic chose to make permission the product. If that framing holds as agents roll out globally, the banks-and-agents standoff that played out in the US this week resolves with less friction than the headlines suggest — because one side built its agent to lose gracefully.
