What I want is a single conversation. Not separate places for research, writing and the news, but one place where I send a message and the work finds its own way from there. AI news, innovation, wine, new ideas: all in the same thread, with someone on the other end who remembers what I asked yesterday.
On 21 October Anthropic announced more or less exactly that. Claude Duet is a separate app in which Claude acts as a proactive agent. You message it the way you’d message a colleague, typed or by voice, and it reaches out to you when something needs your attention.
I just can’t use it. Duet is a research preview for Pro and Max users in the United States, with access through a waitlist. From the Netherlands there’s no way in. So I’m looking under the bonnet and rebuilding it from parts I already have. That’s possible because, since Cowork moved to the cloud, Claude has had scheduled tasks that run without me.
What Claude Duet is
Duet runs on your existing Claude account. It brings over your memory, skills and connected apps, so there’s nothing to set up again. What changes is who starts the conversation. In the regular Claude app you ask and it answers; Duet also messages you.
| Part | What it does | Where it stops |
|---|---|---|
| Main thread | One ongoing conversation, typed or spoken | You can’t pick a model; it runs on Claude Opus 5.5 |
| Channels | Separate threads per topic, sharing the same memory | Anthropic mentions no specialists per thread |
| Morning brief | Looks through your email and calendar each morning and picks what comes first | Only what you’ve connected |
| Its own computer | A virtual machine with a browser that keeps working when your phone is off | You can’t watch it or take over the browser |
| Check-ins | Looks at your accounts at intervals and gets in touch when something needs you | Every few hours, so something urgent can slip in between |
| Tasks and Make | Running jobs in one tab, finished work (research brief, financial model, deck, dashboard) in another | Only you can mark a task as done |
| Approval | Reads and prepares on its own; asks before sending, paying, signing or deleting | Some actions are off-limits even with approval, such as trading securities or moving money between accounts |
Anthropic aims it at professionals and small business owners with more on their plate than they can keep track of. The examples in the announcement are unpaid invoices in QuickBooks or Stripe, missed messages in Gmail and Slack, and payroll in Gusto. That’s a different audience from Claude Code, but the safety layer comes from there: actions pass the same classifier that decides in Claude Code whether a step is allowed without you asking for it.
Proactive means a timetable
Duet doesn’t watch all day. According to Anthropic, Duet checks your connected accounts every few hours, because checking nonstop would burn through your plan’s limits much faster. So proactive here means an agent with a schedule, a memory and the judgement to decide whether something deserves a message.
Anthropic itself names that judgement as what makes Duet useful. The schedule I can build myself. My morning scan has been running every day without me for some time and puts forward three suggestions at most. What I lack is one place where all those separate jobs come together and where I can talk back.
My wish was “every stream in one chat, with sub-agents for each task”. There are two different ideas in that.
- Streams are topics: AI news, wine, innovation. Duet handles those with channels, separate threads that share one memory.
- Sub-agents are specialists, each with its own brief, its own context and its own tools. A researcher who reads sources, a writer who knows my voice, a reviewer who looks at the result with fresh eyes. Anthropic doesn’t mention them for Duet. It decides for itself how much effort a job takes and which apps it needs; nothing in the documentation describes a way to define specialists of your own.
Claude Code does have them. A sub-agent there is a file with a description, instructions and a list of tools. The main agent hands work to it, it does that work in its own context, and only the result comes back. I already use this: a separate reviewer reads my articles once they’re finished, precisely because it didn’t write them. I wrote about that idea in the piece on letting a second agent break things.
How I’m rebuilding Duet
The plan is to lay Duet’s parts next to what I already have and decide part by part. This is what I’ll be testing over the coming weeks.
| Duet | My version | What’s missing |
|---|---|---|
| One main thread | One standing Claude Code session that I steer from my phone with Dispatch and Remote Control | My computer has to be on |
| Channels per topic | A sub-agent per stream: research, writing, AI news, innovation, wine | No shared memory across separate sessions, though there is within the main one |
| Morning brief | My existing morning scan, extended with wine and innovation, posting its results into the main session | Email and calendar aren’t in it |
| Check-ins | Scheduled tasks at fixed times | No judgement about when something is worth a message; I write that down as a rule myself |
| Approval | Permission mode plus my own gates: nothing goes live without my say-so | Already there |
| Make tab | Files in the repo and artifacts | No dedicated view, but version control |
The sixth sub-agent is the one I’m most curious about, and Duet has nothing like it. For now I’m calling it the inventor. Its job isn’t to do work but to look at the work I do and suggest what could be easier. Where do I repeat myself, which step am I still doing by hand, which feature released last week fits something I already do? It may suggest, not build. Building is for the main session, after I’ve said yes.
A plain scheduled task per topic would put me back with separate tabs. The point of Duet is a single point of contact that knows what’s going on. The main session is that point of contact; the sub-agents are its colleagues, and I don’t need to talk to them directly.
What I’ll measure
A home-made Duet is only worth having if it saves time rather than just sending messages. I’m watching four things.
- Messages per day. Anthropic itself says Duet can still misjudge what’s urgent. The same goes for my version. If I get more notifications than I read, it has failed.
- What it costs. Duet counts towards your usage limits too, background work included. Claude Code shows per recurring task how many tokens it uses. I want to set that number against what it delivers.
- Whether the inventor suggests something I actually adopt. If nothing useful comes out, the brief is wrong.
- Whether I still reach for separate tabs. If I do, one main thread isn’t an improvement for me.
The caveat
- The default shifts from asking to being told. An agent that comes to you helps decide where your attention goes. That’s the point, and also the risk: whoever lets something else set the priorities sees less of what wasn’t picked. So I want to write the rules for when something deserves a message myself, rather than leave that to the model’s judgement.
- Approval isn’t the same as oversight. Duet asks before every action with consequences. That sounds safe, but with enough requests a day, approving becomes a reflex. Tap “yes” without reading and you still have control on paper, not in practice.
- Background work eats into your limits. The morning brief and check-ins count towards your five-hour and weekly limits at the same rate as ordinary chats. Anthropic warns you may hit those limits sooner than with chat alone. Being proactive isn’t free, even on a subscription.
- More flows out than your own data. An agent that reads your email also reads email from people who never agreed to anything with Anthropic. Anthropic describes a dedicated classifier meant to stop Duet passing on sensitive information. Still, the Help Center notes that memory keeps what Duet picked up from an app earlier, even after you disconnect that app.
What I’m doing myself
I’m waiting for Duet to reach the Netherlands and building the version above in Claude Code in the meantime, with the six sub-agents as the first test. I’ll write up what it delivers later, even if the answer is that one main thread doesn’t work for me.
If Duet comes to the Netherlands, I’ll run the two side by side. What will count for me is which of the two asks less of my attention.
Frequently asked questions
Can I use Claude Duet outside the US?
Not yet. During the research preview Duet is only available in the United States, for Pro and Max subscribers, through a waitlist. Team and Enterprise plans don’t have it either. Anthropic hasn’t given a date for other countries.
How much does Claude Duet cost?
Nothing extra during the preview: it’s included in Pro and Max. It does count towards your usage limits at the same rate as ordinary chats, and its background work counts too. Anyone admitted gets one full reset of those limits, usable until 30 November 2026.
Does Claude Duet have sub-agents?
Anthropic doesn’t mention any. Duet works with one main thread and separate channels per topic that share the same memory. Specialists with their own brief are something you build in Claude Code, as sub-agents.
How is it different from Claude Code?
Duet is built for knowledge work across your apps, such as email, calendar and accounting, and runs on its own computer in the cloud. Claude Code runs on your own machine or in a cloud session, with your files and your configuration, and lets you decide which agents and rules exist.
Sources
- Introducing Claude Duet — Anthropic, 21 October 2026
- Get started with Claude Duet — Claude Help Center, 23 October 2026
- Claude Duet Safety, Privacy, and Security — Anthropic, 22 October 2026
- Anthropic Launches Claude Duet, a Proactive AI Agent — MacRumors, 21 October 2026
Checked on 25 October 2026 against Anthropic’s announcement, Help Center article and safety post. I haven’t been able to use Duet, so everything about how it behaves in practice comes from those sources rather than my own experience. The three Anthropic pages don’t contradict each other. Whether Anthropic rules out custom sub-agents in Duet or simply doesn’t mention them, I couldn’t establish.