From Chatbot to Coworker That Speaks First
For three years, using an AI assistant has meant the same thing: you open a chat, you type, it answers, it stops. Nothing happens until you ask.
Claude Duet changes that default. Anthropic describes it as an app where Claude "works as a proactive agent." It connects to the tools you already use, keeps an eye on them, and reaches out when something needs your attention. You can still give it tasks, but the point is that it does not wait for you.
It targets a specific audience: professionals and small business owners who have more on their plate than they can track. Think of the owner of a 20-person company who handles invoices, hiring, customer emails and marketing alone.
What Duet Does Day to Day
A Morning Brief
Each morning, Duet sends a short brief on the most important items across your connected apps: the unanswered client email, the invoice that is overdue, the meeting that moved.
Monitoring, Then Silence
During the day, it checks your connected accounts every few hours and messages you only when something needs attention. Anthropic says it spaces out these checks on purpose, to save your usage limits.
Real Tasks, Small and Large
You can hand it work in one ongoing thread, by text or by voice. Anthropic's examples range from simple admin to substantial projects:
- Scheduling meetings and finding missed messages in Gmail and Slack.
- Checking QuickBooks or Stripe for unpaid invoices and drafting follow-ups from HubSpot data.
- Running payroll in Gusto and sending agreements through Docusign.
- Posting approved job listings on LinkedIn and building social campaigns in Canva.
- Managing a Shopify store (a visual storefront editor is "coming soon").
- Building dashboards across several tools, financial or sales forecasts, prototypes and even web apps.
It Has Its Own Computer
Duet runs on a virtual machine in the cloud, with its own browser. Your laptop or phone does not need to stay on, and the agent does not need access to your local machine. You close the app, and the work continues. It also reuses the memory, skills and connected apps of your Claude account, so there is little setup if you already use Claude.
The Approval Model: Works First, Asks Second
An agent that acts while you sleep is only useful if you can trust where it stops. Anthropic splits actions into three tiers:
| Type of action | Examples | What Duet does |
|---|---|---|
| Prepare | Reading, drafting, analyzing, building a deck or a forecast | Does it on its own |
| Act | Sending a message, making a payment, signing a document, deleting anything | Shows the planned action and waits for your approval |
| Forbidden | Trading securities, moving money between accounts | Refuses, even with approval |
Purchases go through Link, Stripe's digital wallet, using a virtual card, or through Shopify's Shop Pay. You can disconnect any app Duet has access to at any time.
Security and Privacy
Anthropic lists several layers of protection. Here are the ones that matter most for a business:
- No passwords in sight. Credentials are handled the same way as in Claude Managed Agents: Duet never directly sees your passwords or sign-in tokens. Sign-in with logins saved in 1Password is "coming soon," with your permission.
- A leak filter. A dedicated classifier and safety system prompt check each message. Requests to share the most sensitive data, such as credit card numbers or government ID, are rejected. Health and personal finance information is shared only when you explicitly ask.
- Prompt injection defenses. Protections against hidden instructions on web pages come from Anthropic's browser-use research.
- Outside testing. Anthropic says internal red teams and an external firm penetration tested Duet, and it is covered by a public bug bounty program.
Anthropic says it will publish more details on its safety design "in the coming days." For a product that reads your email and your accounting, that document is worth reading before you connect everything.
The Fine Print
Usage Limits
There is no extra price: Duet is included in Pro and Max plans. But it shares your plan's usage limits with the rest of Claude, and it uses them differently from a chat. Background checks count, and large tasks like a financial model or a deck take many model turns. Heavy users should watch their limits in the first weeks.
Known Weak Spots
- It can misread a thread or get wrong what is urgent. Anthropic says so itself.
- Personal errands, like booking tickets, are not its strength. The focus is work.
- It cannot yet show you its browser while it works, so you see results, not the clicks.
How It Compares
Duet arrives in a crowded field. TechCrunch lists Meta Muse, OpenAI Dots, Grok Bot, Hark and Instinct among the personal and productivity agents it competes with. Meta's Muse is the one to beat on reach: Meta says it passed 10 million US downloads in 40 days, with shopping as a top activity.
Anthropic's bet is different. Duet is not pitched as a consumer assistant that books your dinner, and unlike Instinct it does not stream its browser activity live. It is pitched as a work agent with a large catalog of business connectors and strict limits on what it may do with money. If you run a small company, that is the right trade-off. If you want an assistant for your personal life, other agents may fit better today.
Who Should Join the Waitlist
Where Duet Makes Sense
- Small business owners: one agent watching invoices, payroll, customer messages and the online store, with a morning summary.
- Consultants and freelancers: client follow-ups, scheduling and weekly reports without a virtual assistant.
- Managers buried in tools: a single brief across Gmail, Slack, the CRM and the calendar.
- Builders: prototypes and small web apps that run in the background while you work on something else.
- Existing Claude Pro or Max users: your memory, skills and connectors carry over, so trying it costs little.
How to Get Access and Start Safely
1. Join the Waitlist
Sign up on the Claude Duet page. You need a Claude Pro or Max plan, and the preview starts in the United States.
2. Connect Little at First
Start with two or three apps, for example your email, your calendar and one business tool. Add more once you trust its judgment. Every connection is access, and access is risk.
3. Read Every Approval
The approval step is your safety net. In the first weeks, read each planned action carefully instead of approving by reflex. This is also the best way to learn how the agent reasons.
4. Measure It
After two weeks, ask yourself three questions: did the brief catch things you would have missed, how much time did it save, and how fast did it use your plan limits? Keep it if the answers are good.
Claude Duet turns Claude from a chatbot into a background coworker. A morning brief, constant monitoring, real tasks on its own cloud computer, and clear rules: it prepares on its own, asks before it acts, and never moves money between accounts. It is a US-only research preview with real limits, but for small businesses it is one of the most practical agent designs so far.
Frequently Asked Questions
My Take
As a network engineer, I spend my days thinking about access: who can reach what, with which rights, and who approves the change. Claude Duet is, at its core, an access design problem.
That is why the most interesting part of this launch is not the model. It is the rules. Prepare freely, act with approval, and never touch certain things at all. That is the principle of least privilege, applied to an AI agent. Add credentials the agent never sees and a filter for sensitive data, and you have a design that a security team can actually reason about.
The risk is human, not technical. After the fiftieth correct suggestion, people stop reading approvals and start clicking "yes." An agent that works while you sleep needs trust that is earned task by task, and checked again after it is earned.
For a small business owner, the pitch is still strong: one assistant that reads your data every morning and tells you what to fix. So, would you let an AI agent run in the background on your business data, or do you want to see every click?
Sources:
- Anthropic: Introducing Claude Duet
- TechCrunch: Claude Duet app is Anthropic's bet on agents
- MacRumors: Anthropic Launches Claude Duet, a Proactive AI Agent
- 9to5Google: Claude Duet is Anthropic's proactive AI agent
- TechCrunch: Meta's Muse hits 10M US downloads in 40 days
This article is published for information purposes only, to help readers discover and evaluate new AI tools.
Related Articles: