Anthropic is putting a new kind of product in front of paying Claude users. Claude Duet, pitched as a “digital chief of staff,” launched as a proactive AI agent for the workplace, according to heise online. Reports describe a research preview that watches your email, calendar and business apps, then comes to you with what needs attention instead of waiting for a prompt.
The timing matters. OpenAI, Meta and Google all have personal or work agents of their own, and Anthropic wants a seat at that table with a product aimed at founders, operators and busy professionals. This article walks through what has been reported about Claude Duet, what Anthropic and its executives have said on the record, how the rivals compare, and what remains unconfirmed as of October 31, 2026.
One caution up front. Most details below come from press coverage and Anthropic’s own product description, and several specifics (the exact launch date, the full list of connected apps, the rival comparison) differ between outlets. Where the record is thin, this piece says so. For background on the model family reportedly behind the product, see our breakdown of Claude Opus, Sonnet and Haiku 5.5.
What Is Claude Duet?
Claude Duet is a new Anthropic app in which Claude works as a proactive agent. That is Anthropic’s own framing: the company said it is opening a research preview of the app. Tatler Asia and other outlets have labelled it a digital chief of staff, a phrase that stuck because it describes the job rather than the technology.
A chief of staff does three things for an executive. They triage the incoming flood, they prepare the briefing, and they push loose ends toward done. Reports say Duet aims at the same loop. It monitors email, calendars and connected business applications, prepares briefings, and drafts or sends messages. Then it keeps going in a virtual machine with its own browser after the user closes the app.
That last point is the real break from the chatbot era. A chatbot answers when asked and goes quiet when you leave. Duet is described as the opposite, an assistant that has a standing to-do list and a place to work while you sleep, commute or sit in meetings.
Reports name Claude Opus 5.5 as the underlying model. Anthropic has not, in the material reviewed for this article, published a technical paper on how Duet decides what counts as urgent. That gap shapes much of the analysis that follows.
When Did Claude Duet Launch? October 21 or 22
The coverage disagrees on the date. Some sources place the research preview launch on October 21, 2026, and others on October 22, 2026. Time zones may explain a one-day gap, since a US evening announcement lands on the next calendar day in Asia and Europe. Chosunbiz, a Korean business outlet, ran its report on October 23 under a headline saying Anthropic launched Claude Duet to challenge Meta and OpenAI in agents.
For readers, the exact day matters less than the status. This is a research preview, not a general release. Anthropic uses that label for products it expects to change quickly, and it usually signals limits on access, features and reliability. Treat any capability list as a snapshot of week one.
This article says “late October 2026” when it refers to the launch and cites the two dates only as reported. If Anthropic publishes a changelog or dated release note, the question will settle itself.
How a Digital Chief of Staff Actually Works
Anthropic’s own description, from its Claude Duet resource page, is the best primary source available. The company says the product “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.” That one sentence holds four design promises, and each one carries risk.
Monitoring and briefings
The first promise is proactive surfacing. Duet reportedly watches your inbox, calendar and business tools, then assembles a briefing of what needs a decision. This is where an agent earns or loses trust. A chief of staff who flags ten false alarms a day gets ignored by day three. One who misses the single overdue contract gets fired. Precision on the “what matters” question will decide whether people keep the app open.
Background work in a virtual machine
The second promise is persistence. Reports say Duet continues work in a virtual machine with its own browser after you leave. That design choice means the agent does not borrow your laptop. It operates from a separate environment, which helps with uptime and also draws a cleaner boundary around what it can touch. It also means every action happens somewhere you cannot see unless the product logs it well.
The third and fourth promises, task completion and check-ins before acting, form the control layer. Anthropic says Duet checks with you before it acts on your behalf. The open question is how often, and on which actions. An approval prompt for every draft would drown users, and an approval prompt for none would scare security teams. Where the dial sits is the product’s most important setting.
Which Apps Does Claude Duet Connect To?
Reported supported services include Gmail, Slack, QuickBooks, Stripe, Gusto and DocuSign. Read that list as a map of a small company’s operations. Gmail and Slack carry communication. QuickBooks and Stripe carry money in and out. Gusto carries payroll. DocuSign carries signed commitments. Together they cover most of what a founder or office manager touches before lunch.
| Service | Business function | Why an agent wants it |
|---|---|---|
| Gmail | Triage, summaries, draft replies | |
| Slack | Team chat | Surface threads that need a decision |
| QuickBooks | Accounting | Spot unpaid or unusual items |
| Stripe | Payments | Track revenue and payment issues |
| Gusto | Payroll and HR | Prepare payroll tasks for review |
| DocuSign | Contracts | Track documents waiting on signatures |
The third column is analysis, not a quote from Anthropic. The reports confirm the integrations, and the use cases are the natural fit for each. Money and signatures are also where mistakes cost the most, so expect the approval gates to concentrate there.
Breadth of connection is the quiet competitive weapon in this category. An agent that sees only email is a smarter inbox. An agent that sees email, ledger, payroll and contracts can notice that a client invoice is late and a payroll run is near, then connect the two. Our earlier look at Google’s connected apps platform for Gemini shows the same land grab from the other direction.
Who Can Use It and What It Costs
Reports say the preview is available to Claude Pro and Max subscribers in the United States through a waitlist. There is no separate Duet price in the reporting reviewed here. Access comes bundled with existing subscriptions, which reportedly cost $20 a month for Pro and $100 or $200 a month for Max.
| Plan | Reported monthly price | Reported Duet access |
|---|---|---|
| Claude Pro | $20 | Waitlist, US only |
| Claude Max (lower tier) | $100 | Waitlist, US only |
| Claude Max (higher tier) | $200 | Waitlist, US only |
| Free tier | $0 | Not reported as eligible |
The waitlist plus US-only limit tells you Anthropic is rationing demand, probably for two reasons. A persistent agent with its own virtual machine costs far more to run than a chat window, and a research preview lets the company watch real behavior before it opens the door wider. Neither point is confirmed by Anthropic. Both fit the pattern of earlier staged launches.
Team and enterprise availability is the real prize, since companies pay more and buy in bulk. Reports do not give a date for that step. A headline from TechTarget, which I cover below, suggests enterprises are not yet ready to receive it anyway.
Claude Opus 5.5 Under the Hood
The reported engine is Claude Opus 5.5, the model tier at the top of Anthropic’s current lineup. Agents lean on a model’s ability to plan across many steps, recover from errors and hold context over long sessions. Raw chat quality matters less. Pairing the top tier with a long-running agent is a logical choice.
Our side-by-side of Gemini 4 Carbon, Claude Opus 5.5 and GPT-6.1 Terra covers how the frontier models differ on benchmarks and cost. For Duet, benchmarks tell only part of the story. A chief-of-staff agent lives or dies on judgment about timing and scope, plus a long list of boring integrations that have to keep working.
There is a practical implication for pricing. Running a top-tier model continuously in the background is expensive. If Duet stays inside Pro and Max subscriptions, usage limits will probably do the cost control, and users should expect caps on how much background work they get. Anthropic has not published those limits in the sources reviewed.
What Anthropic Executives and Staff Are Saying
Five on-record statements are available, and they split into two groups. The company’s own words describe the product. Individual people describe how it feels to use. All five come from the verified quote file for this story.
Anthropic’s launch statement on its Claude Duet resource page reads: “We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent.” The phrase “research preview” sets expectations. The company is telling users the product will shift under their feet.
The same page offers the sentence quoted earlier about coming to you with what needs attention. It is the closest thing to a specification the public has, and it contains the promise to check with you before acting.
Mike Krieger, co-lead of Anthropic Labs, spoke to Business Insider about judgment. He said Duet “can have the discernment and judgment around when to be proactive, when to act on your behalf.” That claim is the heart of the product. Proactivity without discernment is spam with a robot voice.
Rahul Patil, Anthropic’s chief technology officer, struck a more candid note in a piece published by State of AI. He said: “Giving an AI the keys to your Slack, mail, calendar and everything else is a little scary. I know, because I did it.” A CTO admitting fear is useful data. It tells buyers the company knows the trust problem is real, and it tells security teams the risk conversation is worth having. Our earlier piece on Anthropic’s leadership bench, including Patil, gives more background on who he is.
The last quote comes from Frank Lee, identified as an Anthropic employee, in a post on X. He wrote that the best part of working with Claude Duet is that “it proactively identifies and pushes along important work i’m slipping on – it’s highkey the best chief of staff agent.” Read it for color, not proof. Lee works at the company, so the endorsement is an inside view, not an independent review.
Claude Duet vs OpenAI, Meta and Google Agents
Reports name competing agent products at all three rivals. OpenAI’s is called Dots. Google’s include Gemini Agent and Gemini Spark. Meta’s belongs to its Muse line. The results reviewed do not establish a complete, verified feature-by-feature comparison, so the table below sticks to what outlets attribute to each company and leaves specs out.
| Company | Reported agent product | What is established |
|---|---|---|
| Anthropic | Claude Duet | Research preview, proactive agent, Pro and Max waitlist in the US |
| OpenAI | Dots | Named in reports as a competing agent, details not verified |
| Gemini Agent, Gemini Spark | Named in reports as competing agents, details not verified | |
| Meta | Muse | Named in reports as the basis for a personal agent, details not verified |
Chosunbiz framed the launch as Anthropic challenging Meta and OpenAI in agents, and Business Insider’s coverage of Krieger’s remarks likewise cast Duet as Anthropic’s answer to the new personal agents from Meta and OpenAI. Both framings agree on the direction of the race. All four labs now believe the next interface is an assistant that acts, not one that chats.
Where Duet may differ
Based on reported design, Duet’s sharpest edge is its business focus. The integration list (accounting, payroll, payments, contracts) reads like a small-company back office. Consumer-leaning agents tend to lead with shopping, travel and family calendars. If Anthropic holds that line, Duet competes for the person who runs a company, not the person who runs a household.
Wired published a piece titled “I Tested Every Major AI Agent for a Day. Here’s the Best One,” which signals hands-on comparisons are arriving fast. Wired’s agent test is worth reading in full, and I have not relied on its verdict here. Readers weighing the products should wait for several independent tests, because a one-day trial cannot measure how an agent behaves in week six.
Why Enterprises Are Not Ready Yet
TechTarget ran a story under the headline that Anthropic rolled out Claude Duet but enterprises aren’t ready. I am relying on that headline alone for the claim, so the following reasoning is my own analysis of why that might be true.
Large companies buy software through procurement, security review and legal sign-off. An agent that reads every email and can message customers hits all three gates at once. The security team asks where the data lives, who can see the agent’s logs, and what happens if someone tricks the agent with a hostile email. Those questions have no standard answers yet, because the category is months old.
The identity question is the sharpest. An agent logs into systems as a user, or as something like one. Most identity setups were built for people, and they strain when software acts with a person’s authority around the clock. Our comparison of six identity providers for AI agents shows how vendors are retooling for exactly this. Secrets handling is the sibling problem, and our coverage of Infisical’s $40 million raise for AI agent security shows investors are betting on it.
Small teams face fewer gates, which fits Anthropic’s decision to start with individual Pro and Max subscribers. A founder can say yes in an afternoon. A bank cannot.
Market Impact: The AI Race Becomes a Race for Access
State of AI argued in a headline that the AI race is no longer about intelligence and now comes down to access. That framing explains why Duet matters beyond its feature list. Once several models reach similar quality, the winner is the one with permission to touch your mail, your calendar and your books. Whoever holds those connections owns the daily habit, and the habit sells the subscription.
For Anthropic, the commercial logic is straightforward. Chat subscriptions face churn because users can switch with a click. An agent that has learned your priorities, your vendors and your writing voice carries switching costs. Each connected app adds a little more stickiness.
For software vendors, the picture is double-edged. Gmail, Slack, QuickBooks, Stripe, Gusto and DocuSign gain a new channel when an agent drives their products. They also risk becoming back-end plumbing that users never open. Interface power may shift from the apps to the agent that sits on top.
Investors will read this launch alongside the broader AI spending debate. We covered that tension in our look at Big Tech’s $750 billion AI bet. Agents are one of the clearest stories the industry can tell about how that spending turns into revenue, because a persistent agent justifies a recurring bill in a way a chat answer does not.
Historical Context: From Chatbots to Agents That Act
Digital assistants are not new. Voice assistants on phones and smart speakers promised to run your day more than a decade ago and mostly delivered timers and weather. They could not read a contract or reconcile a ledger, so they stayed small.
The chatbot wave changed the intelligence side of the equation. Models could suddenly write, summarize and reason. Yet they stayed passive. You pasted text in and copied text out. The next step, coding agents that edit files and run commands, proved that models could act inside a bounded environment with a clear goal. Anthropic’s own coding tool grew out of that shift, and we have tracked its quirks in our report on Claude Code’s game-recreation block.
Duet takes that coding-agent pattern and points it at office work. The environment is messier, the goals are vaguer and the stakes include payroll and legal documents. If coding agents succeeded because tests tell them when they are wrong, chief-of-staff agents must succeed without a test suite. That is a harder problem, and it is why the human check-in stays central.
What Is Still Unconfirmed About Claude Duet
Plenty of the story rests on reports instead of verified specifics. Here is where the record stays open.
- Exact launch date. Sources split between October 21 and October 22, 2026.
- Underlying model. Claude Opus 5.5 is reported, but Anthropic’s own material reviewed here does not spell out the configuration.
- Full integration list. Gmail, Slack, QuickBooks, Stripe, Gusto and DocuSign are reported, and other services may be supported or planned.
- Usage limits. No verified caps on background work or task counts.
- Approval rules. Anthropic says Duet checks before acting, but the exact actions that trigger a prompt are not documented in the sources reviewed.
- Rival comparison. Dots, Gemini Agent, Gemini Spark and Muse are named, though no verified side-by-side exists.
- Team and enterprise timing. No confirmed date.
None of these gaps is unusual for a research preview. They do mean buyers should treat feature lists as provisional and test with low-stakes accounts first.
Five Predictions for AI Chief-of-Staff Agents
These are my forecasts, not reported facts. Each is a judgment call based on the pattern of earlier AI launches.
- The waitlist widens within months. Research previews at AI labs usually expand to more users and regions fairly quickly once capacity and early feedback allow. Expect Duet to reach more Pro and Max users outside the US first, then teams.
- A security incident involving an agent will make headlines. Prompt injection through a hostile email is the obvious path. The first public case will shape enterprise policy for the whole category.
- Approval controls become a selling point. Vendors will compete on granular permissions, audit logs and per-action settings. The product with the best dial will win cautious buyers.
- Rivals copy the chief-of-staff framing. OpenAI, Google and Meta already ship agents, and marketing will converge on the same job-title language, with business tools at the center.
- Integration depth beats model rank. Once top models cluster in quality, the agent with the most reliable connectors will pull ahead. More of our coverage on this trend lives in the AI and machine learning section.
The takeaway: Duet is less a model release than a bet on trust. Anthropic is asking users to hand over the keys, and its CTO openly admits that feels scary. If the company earns that trust through restraint, logs and clear approvals, the product has a shot. If not, rivals will take the same opening.
FAQ: Claude Duet
What is Claude Duet? Claude Duet is an Anthropic app in research preview where Claude works as a proactive agent. Reports call it a digital chief of staff that monitors email, calendars and business apps, prepares briefings and keeps working after you leave.
When did Claude Duet launch? Coverage says the research preview began on October 21 or October 22, 2026. The sources disagree, so treat it as late October 2026.
Who can use Claude Duet and what does it cost? Reports say Claude Pro and Max subscribers in the United States can join a waitlist. Reported plan prices are $20 a month for Pro and $100 or $200 a month for Max. No separate Duet fee has been reported.
Which apps does Claude Duet work with? Reported supported services include Gmail, Slack, QuickBooks, Stripe, Gusto and DocuSign. Anthropic has not published a final list in the material reviewed.
Does Claude Duet act without asking? Anthropic says Duet checks with you before it acts on your behalf. The exact actions that trigger a prompt are not documented in the sources reviewed, so test with low-risk tasks first.
How does Claude Duet compare with OpenAI, Meta and Google agents? Reports name Dots at OpenAI, Gemini Agent and Gemini Spark at Google, and Muse at Meta as competing products. No verified feature-by-feature comparison exists yet, and independent hands-on tests are still arriving.
Is Claude Duet ready for enterprises? A TechTarget headline says enterprises are not ready. Reports do not give a date for team or enterprise availability, and security, identity and audit questions remain open.