The Black Cloud

AI - Anthropic introduces Claude Duet, a proactive personal assistant

I do not like that I have to do Anthropic news twice in a row, but this one is definitely worth a blog 🤣

Anthropic is joining the always-on personal agent race with Claude Duet. Rather than waiting for you to open a chat and think of the right prompt, Duet is meant to keep an eye on your work and come to you when something needs attention. That sounds useful… and also like a product where trust matters a lot more than a flashy demo 😊

However, with all the competition from Muse, Dots, Grok Bot and others, it will be interesting to see how Duet differentiates itself and earns users’ trust.

So far it has quite a limitation that I do not like; it is not a chief of staff yet because it can see the Claude sessions on my computers but cannot direct them. So not sure how useful it would actually be.

Still the war for always-on personal agents is waging, and I am yet to pay for one myself, kind of waiting to see which one truly wins 🤣

Also with Grok Bot also using Opus 5.5, so far it seems probably like a better option? Depending for what you really need it… For personal use I guess Grok Bot wins, but if you wanted to offload some of the actual work work 😉 to your phone agent - this could be where Claude Duet shines.

Claude Duet

Claude Duet is a separate app for iOS and Android, and is also available on the web at claude.ai. It is powered by Claude Opus 5.5 and uses the memory, skills and connected apps from your Claude account. You can message it by text or voice in one ongoing thread, while it works in the background on its own cloud computer with a browser.

The idea is that Duet sends you a morning brief, checks your connected accounts every few hours and only interrupts you when it thinks something needs your attention. You can also hand it work: Anthropic’s examples range from finding emails and drafting replies to scheduling meetings, preparing payroll, building dashboards and creating forecasts. It can use more than 6,000 connectors and plugins from the Claude Marketplace.

It’s not just a chatbot with a longer to-do list. The interesting part is the hand-off: Duet is meant to keep going while you’re away, then notify you when the work is done or it needs a decision. You still approve consequential actions. It can prepare a message, payment, signature or deletion, but must show you what it intends to do and wait for your approval before carrying it out. Some actions, including trading securities and moving money between accounts, are off-limits.

Anthropic also says Duet can buy items for your business through Stripe’s Link, with your confirmation and a virtual card; it can check out with Shop Pay too. Logging in with saved 1Password credentials is planned for later.

NOTE

Claude Duet is in research preview and access is through a waitlist for Claude Pro and Max subscribers in the US. It is not available to Team or Enterprise subscribers in this preview.

Duet shares your Claude plan’s usage limits. Its background checks count towards those limits, and bigger jobs can use more model turns and background work than a regular chat. Anthropic also says this is an early version: it can misread a thread or get the urgency wrong, personal errands are not its strong suit yet, and it cannot currently show you its browser.

Official announcement

We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent.

It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.

Anthropic’s Introducing Claude Duet explains how the preview works and what it can do.

The hard part is judgement

The useful promise here is not simply “Claude can do tasks”. It’s that it can decide what to bring to you, what to work on in the background and when to stop for permission. Anthropic says Duet does not directly see your passwords or sign-in tokens, and that it has protections against prompt injection and the accidental sharing of sensitive information. Those are sensible design choices, but they are not a guarantee that an agent will always understand the situation correctly. The preview warning about urgency is worth taking seriously.

This is the same bigger question facing every proactive agent: how do you make it helpful without having it interrupt you constantly, and how do you make it act without giving it too much authority? A careful approval step is important, but the quality of the agent’s judgement before it asks is just as important.

What people are saying

Anthropic’s Amol Avasare says Duet is designed for work and has been the main way he gets work done during the past few weeks:

Claude Duet is out today in research preview!

Over the past few weeks, it’s become the main way I get work done.

Lots of personal agents these days, but this one is built for work, has very high judgement, and indexes heavily on safety / security.

That is a strong endorsement from someone at Anthropic, so I’d treat it as the company’s perspective rather than independent evidence. Early testers have also highlighted how Duet turns one conversation into multiple tracked tasks. Alex Finn describes that workflow in his first impressions:

Dan Shipper says they have been testing Duet for two weeks. He likes the dedicated app, task tracking and document creation, but also points out that it is not yet a full “chief of staff”: it cannot direct his Claude sessions, and its voice experience is transcription rather than a spoken conversation.

And a useful counterpoint from elvis: personal agents may not all collapse into one winner. People will care about trust, personality and how well an agent understands their intent, and may choose different assistants for different kinds of work.

My take

I’m interested in the shape of this: one place to hand off work, with background progress and a clear approval step, is much more appealing than babysitting a collection of chat windows. The work focus and connection to Claude’s existing memory and apps could make it genuinely handy.

But “always on” only works if the interruptions are rare and well judged, and a connected agent needs to earn trust task by task. This is still an early, US-only research preview with usage coming from the same limits as regular Claude, so I wouldn’t take the launch hype as proof that the hard parts are solved. Still, I’m very curious to see how Duet handles the ordinary, messy stuff that doesn’t fit neatly into a demo 😊

<< Previous Post

|

Next Post >>

#Blog #AI #Claude #Anthropic #Personal AI #Agents #Claude Duet