Claude Duet Explained: Your AI Assistant Wants a Job, Not Another Prompt
Anthropic's new Claude Duet promises to work in the background, watch connected apps and bring important work to you. It also raises a more awkward question: how much supervision does an always-on assistant need?
Lynka Team · October 28, 2026 · 10 min read · AI
Follow Lynka on Google
Get new articles in your preferred sources.
On this page
There is a particular kind of software demo that always looks fantastic.
An AI opens your inbox, notices three important things, moves a project forward, prepares a report, and quietly leaves you alone.
That is roughly the promise behind Claude Duet, Anthropic's new proactive AI agent.
It was announced on October 21, 2026, and the pitch is not subtle: instead of waiting for prompts, Claude can keep working between conversations and bring things to your attention. Anthropic is starting with a research preview available through a waitlist for US Pro and Max subscribers. Anthropic's announcement explains the launch and current availability.
This is an interesting release. Not because Claude can now draft another email.
Because the product is asking to move from answering questions to holding responsibilities.
The catch is that a responsibility never comes alone. It brings permissions, mistakes, supervision, cost, and one uncomfortable question:
How much of your work are you willing to let an AI touch when you're not there?
What is Claude Duet?
Claude Duet is a proactive AI assistant built around Claude Opus 5.5. It is intended to notice what matters in your connected work apps, complete tasks you assign, and continue operating after you step away.
Anthropic describes a morning briefing that pulls important items from connected tools, followed by background checks throughout the day. You can also assign work that takes longer than a chat reply, such as preparing a forecast, building a dashboard, or handling administrative tasks. Anthropic describes those capabilities in the launch post.
That is different from opening a chatbot and asking it to summarize your day.
A regular chatbot starts when you ask.
A proactive agent is supposed to notice that the day has changed.
The distinction sounds small until you consider a normal business owner with 14 unread emails, three customer complaints, a payment waiting for confirmation, and a proposal nobody followed up on.
The hard part is not writing a reply to any single message.
The hard part is knowing which message deserves attention before it becomes a problem.
What Claude Duet can actually do
Anthropic lists examples across connected business tools, including reading Gmail and Slack messages, checking finance systems, drafting reminders, preparing documents, working with CRM records, and building or updating business materials.
It also describes connected-app workflows through the Claude Marketplace, including tools such as HubSpot, QuickBooks, Stripe, Canva and Shopify. These are examples of supported workflows, not a promise that every account can use every integration without setup or approval. See Anthropic's supported workflow examples.
Three types of task stand out.
Tasks you know you need
Prepare a summary of the outstanding invoices and draft reminders for review.
You know the task exists. You simply don't want to open four systems to complete it.
Things you might have forgotten
Check whether tomorrow's meetings have missing preparation documents or conflicting times.
This is more interesting. The agent must recognize a problem before you ask a question about it.
Work that develops over time
Keep track of the three proposals awaiting responses and tell me if something materially changes.
Now the value is continuity. The agent needs to remember what it was watching, not restart research from scratch every morning.
Who can use Claude Duet right now?
As of October 26, 2026, Claude Duet is a research preview, not a product everyone can activate immediately.
Anthropic says it is starting with users in the United States on Pro and Max plans, via a waitlist. The app is available on iOS, Android and the web for eligible participants. It lives separately from the main Claude app during this preview. Anthropic's getting-started guide covers access.
Do not read a launch headline and assume the feature will be waiting in your account settings today.
Early access is early access.
Is Claude Duet free?
Not as a standalone free product at launch.
Anthropic says the preview is included with eligible Pro and Max subscriptions at no extra charge. But its work consumes the same plan usage allowances, including background checks and larger multi-step tasks. The company also says admitted preview users receive a one-time usage limit reset, subject to its stated window. See the official usage guidance.
That difference matters.
Included in your plan is not the same as unlimited work.
A morning briefing that sounds free can still consume part of the allowance you wanted to use for something else.
Claude Duet vs a scheduled automation
It is worth being skeptical of any feature labelled "proactive."
A scheduled automation can already say:
Every morning at 8, send me today's calendar.
You do not need a frontier model to know that 8 AM exists.
Claude Duet is more interesting when the task requires judgment:
Review the calendar, recent messages and active projects. Tell me which three things need an actual decision today, and leave the routine stuff alone.
The key difference is not the clock.
It is the decision about relevance.
Of course, that means the agent can decide badly. It may miss something important or confidently prioritize a message that could have waited until Friday.
That is why the product should be evaluated on the quality of its judgment, not the number of tasks shown in a launch video.
Claude Duet vs OpenAI Dots
Anthropic isn't the only company pursuing persistent assistants. OpenAI introduced Dots, its own approach to ongoing AI work, earlier this autumn.
Both products move away from the one-prompt-at-a-time model. Both are designed around retained context and work continuing beyond the immediate conversation.
What is actually worth comparing?
Question
Claude Duet
OpenAI Dots
Basic idea
Proactive Claude assistant
Persistent ChatGPT agent
Important distinction
Research preview with explicit US Pro/Max waitlist at launch
Availability and capabilities depend on ChatGPT plan and rollout
Work context
Connected tools and ongoing work
Connected tools and persistent context
Risk to evaluate
Background usage and approval controls
Background usage and approval controls
Real test
Does it surface the right issues?
Does it surface the right issues?
Do not choose either product because one has a friendlier mascot.
Give both the same bounded job, where available, and measure how many times you have to rescue them.
The real competition is over your attention
Software used to compete for screen time.
Open the CRM. Open email. Open the project board. Open the invoice tool.
Persistent agents want to do something more ambitious: decide when you need to open any of them at all.
That can be excellent.
Imagine the agent quietly handles the obvious things and only interrupts when a customer is about to churn or a deal needs a decision.
Now imagine it tells you every time someone edits a document title.
Congratulations. You built Slack with more expensive notifications.
The best proactive AI may be the one that knows when to keep quiet.
That is a harder product problem than adding browser access.
What about security and privacy?
A proactive agent gets useful by seeing more context. That is exactly what makes it sensitive.
Anthropic says Claude Duet uses an isolated credential-handling design, automated action checks, and additional screening for sensitive information sent in messages. It also describes defenses against prompt injection, external testing and approval steps before consequential actions. Anthropic explains its safety architecture.
Its help documentation says Duet seeks approval for actions such as sending messages, payments, signing documents or deleting items.
Those are safeguards, not proof that the system can never make a mistake.
The dangerous scenario is not necessarily an evil AI.
It is a perfectly normal assistant reading a misleading email, making an incorrect assumption, and trying to do a helpful thing in the wrong account.
What I would not connect on day one
I would avoid giving a new agent broad, unnecessary access to sensitive financial systems, production infrastructure, or administrative accounts.
Start with read-only research and drafts.
Then earn trust through actual tests.
What I would connect first
For a business owner, a sensible initial scope might be:
calendar and selected project context
email summaries rather than automatic sending
CRM opportunity read access
internal task lists
a place to save drafts
You can learn a lot without granting the ability to delete a customer or commit company funds.
Five business tasks worth testing
1. The overdue follow-up brief
Ask the agent to find leads or deals with no recent activity, explain why they matter, and propose next steps.
Judge it on relevance and false alarms.
2. Invoice follow-up preparation
Ask it to identify open invoices and draft reminders, without sending them.
Check amounts and payment status manually against the source system.
3. Pre-meeting research
Have it summarize the customer history, outstanding issues and proposed agenda before a sales meeting.
This is high-value work with relatively limited operational risk.
4. Weekly competitor changes
Define five competitors and three types of change worth reporting.
The test is whether the agent filters noise, not whether it finds every homepage update.
5. Project handoff summaries
Ask it to explain what changed since the last discussion and what still needs a decision.
The key is context retention. If you must re-explain the project every time, persistence has not earned its keep.
How to judge whether Claude Duet is useful
Do not measure how busy it looks.
Measure outcomes:
Metric
Why it matters
Important issues correctly surfaced
Did it notice what you would have missed?
False alarms
How often did it interrupt for nothing?
Human review time
Did you actually save time?
Mistakes caught before action
Are approvals doing useful work?
Useful tasks completed
Did anything get finished?
Usage consumed
Was the result worth the cost or allowance?
A flashy agent that creates 20 reminders and requires 35 minutes of checking is not necessarily better than a simple daily checklist.
What this means for CRM software
The interesting part for sales teams is not an AI writing emails. That was already possible.
It is an assistant noticing that a qualified opportunity has stalled, that a customer has an unresolved complaint, or that a proposal has not received a response.
But the CRM still needs accurate records.
If the deal owner is wrong, the date is stale, and the activity history is missing, your proactive assistant will faithfully automate confusion.
That is why Lynka's underlying workflow remains important: leads and deals, lead generation, and clear next actions. The agent might become the interface, but the record underneath still has to make sense.
The actual question
Will Claude Duet be useful?
Possibly very.
Will it be useful simply because it keeps working when you close the app?
Not necessarily.
A bad process running continuously is still a bad process.
The prize is an assistant that remembers enough, judges well enough, and interrupts rarely enough that you can stop supervising every small thing.
Claude Duet is Anthropic's proactive AI agent for ongoing work across connected tools and tasks. It was announced on October 21, 2026.
No. The initial research preview uses a waitlist and begins with US Pro and Max users.
During the preview, Anthropic says Duet is included with eligible Pro and Max subscriptions without an extra fee, but its actions and background checks count toward existing usage limits.
It is designed for background work and proactive updates. Actual access and permitted actions depend on preview eligibility, connected apps and approvals.
Anthropic describes approval requirements for consequential actions including sending messages. Review the current settings and official documentation before trusting a particular workflow.
There is no defensible universal winner. Compare the same real task, source access, failure rate, human supervision and ongoing cost.
