Anthropic opened the waitlist for Claude Duet on October 21, and the first thing worth knowing about it is in the fine print, not the demo. Duet’s product page says the agent “can miss something that arrives between checking in with you.” That one sentence tells you more about how it works than the 6,000 connectors on the same page.
Duet is the third always-on personal agent from a major lab in seven weeks. Meta shipped Muse on September 8, OpenAI shipped dots at DevDay on September 29, and now Anthropic has its answer, powered by Opus 5.5 and limited for now to US subscribers on the Pro and Max plans. MacRumors framed it as Anthropic’s response to Muse and dots, which is fair. All three run on a cloud computer of their own with a browser, all three connect to thousands of apps, and all three ask before they do something sensitive. If you read the launch coverage side by side, they blur together.
They are not the same product. The differences sit in three places most coverage skipped: what each agent does when you haven’t asked it anything, who pays for that background work, and what the vendor can see.
The three agents, compared on what they do unprompted
| Claude Duet | OpenAI dots | Meta Muse | |
|---|---|---|---|
| Launched | Oct 21, 2026 (waitlist) | Sep 29, 2026 | Sep 8, 2026 |
| Model | Opus 5.5 | GPT-6 Astra | Meta’s Muse models |
| Who can get it | US Pro and Max subscribers; no Team or Enterprise yet | Pro and Business Premium; Pro excluded in EEA, Switzerland and UK | US adults 18+ |
| Entry price | Claude Pro, $20/month | ChatGPT Pro | Free tier; paid tiers reported at $20 and $100 |
| Background behavior | Morning brief, checks apps every few hours, drafts next steps | Read-only “proactive research” when idle | Keeps working after you close the app |
| Stated approval gates | Sending messages, payments, signatures, deletions | Custom allow/block/approve rules plus auto-review; password changes always stay with you | Purchases and email |
| Who pays for background work | Counts against your plan limits | Conversations don’t count; Codex and Work tasks do | Token allowance by tier |
Sources for the table: Anthropic’s Duet page and pricing page; The Next Web on dots; The Neuron on Muse, which notes that the Muse plan names and prices came from TechCrunch reporting and could not be confirmed on a live pricing page. The approval-gate rows are the vendors’ stated policies, not behavior anyone outside the labs has measured.
The row that separates them is background behavior. OpenAI made the idle state of a dot structurally read-only: it can look at your connected apps but cannot send, change or click anything until you hand it a task. Duet goes further on its own. It reads, prioritizes and drafts, then, per the product page, “goes quiet until something needs you.” Muse is built around finishing long tasks after you leave. Same architecture, three different answers to the question “what is this thing allowed to do at 3 a.m.?”
Duet is a poller, and that sets what it is good for
The miss-between-check-ins caveat is not modesty. It describes the design. MacRumors reports that Duet “checks apps every few hours” after the morning brief. It is not subscribed to your inbox; it visits it on a schedule.
I spent a lot of years in IT operations at a large telecom arguing about exactly this distinction, polling versus traps. A monitoring system that polls a router every five minutes is fine for capacity graphs and useless for telling you a link dropped four minutes ago. Nobody put an SLA on a poller. We used polling for trends and traps for anything that paged a human. A personal agent that looks at your apps every few hours belongs in the first category: a strong morning brief, invoice chasing, contract prep, the weekly sweep. It is the wrong tool for anything where a two-hour delay costs money, like a customer escalation or a payment that has to clear today. Anthropic says so on its own page; the useful move is believing it.
The polling design also explains the metering. Every check-in is model work, and Anthropic states that Duet’s background checks count toward the same plan limits as the rest of Claude, with larger tasks consuming more turns. On a $20 Pro plan, an agent that wakes up every few hours across a dozen connected apps is spending your allowance while you sleep. If you already hit Pro limits in Claude Code or chat, Duet will compete with that work. OpenAI made the opposite choice for dot conversations, which is one of the few places the products differ on cost rather than capability.
What the vendor can see, and what Duet hasn’t said yet
On credentials, the three are close. Duet does not see passwords or logins, and Anthropic says approvals, logins and safety checks run in your own cloud computer. Dots sign in with saved passwords without exposing them to the model. Muse keeps credentials apart from the agent through a separate permissions system Meta calls Sentinel.
On data, the gaps show. The Neuron notes that Muse’s launch version “does not prevent Meta from accessing data when necessary to operate, support, or secure the service,” with a confidential VM promised later. Duet’s page says sensitive details “stay private” but gives no retention period, no storage detail and no statement on training use. SiliconANGLE reports that an automatic classifier screens what Duet sends for sensitive data leaks, and that the agent sits under the same trustworthy agents framework Anthropic uses for Claude Code. That covers outbound actions. It does not answer what Anthropic keeps from a cloud computer that has read your QuickBooks and your payroll for a month.
That matters because Duet is pitched at professionals and small business owners, the exact users whose connected apps hold client and employee data. The same logic we applied to dots holds here: each app you connect is a credential grant, and the connect decision sets your exposure far more than the model does.
Which one, if you’re choosing this month
If your work already lives in Claude (memory, skills, connected apps carry straight over) and what you want is a daily triage layer, join the Duet waitlist, connect read-heavy apps first, and watch how fast it eats your plan limits for a week before you connect anything with a payment button. If you want an agent that does nothing on its own until asked, dots’ read-only idle mode is the more conservative default, provided you are not a Pro user in Europe or the UK. If you want long unattended tasks like booking and buying, Muse is built for that, and its collision with retailers like Amazon is the part to watch.
For any of the three, keep your own record of what the agent did rather than relying on its summary; the case for an independent trace applies to a personal agent touching your books as much as to a production one. Duet is a research preview, and Anthropic says plainly that it “may misread a thread or miscalibrate what’s urgent.” Treat its morning brief as a strong first draft of your day, not as the system of record for it.