Anthropic Launches Claude Duet: the "Agent That Finds You" Enters Research Preview

Drawing the line: this is "the third one," not "a Dots clone"

One line to draw first, so nobody misreads this. After OpenAI Dots and Meta Muse, Anthropic's Claude Duet is the third product to charge into the "proactive agent" race — all three agree that "the agent comes to you" is the product paradigm of Q4 2026. But our earlier Dots piece covered OpenAI's proactive-agent play; Duet is a different vendor and a different product line: a different model underneath, a different permission design, a different target user, with zero overlap in events and facts. If you have never been interrupted by an "agent that finds you," this piece maps the whole new species for you; if you already use Dots or Muse, Duet's wedge — "the daily chores of a small-business owner" — is another answer worth comparing.

The one-sentence fact: on October 21, 2026, Anthropic announced that Claude Duet opens as a research preview, by waitlist in waves, for Claude Pro and Max subscribers in the US only; Team and Enterprise are not supported yet. It is not a toggle inside the main Claude app but a standalone app: its own iOS and Android apps, plus claude.ai on the web. A note of honesty on sourcing: the original @claudeai post could not be opened during scouting (403), so for this piece I opened Anthropic's official announcement page myself (https://claude.com/resources/articles/claude-duet), whose footer reads "Published: 2026-10-21" — fully consistent with the six independent outlets (MacRumors, 9to5mac, The Verge, and others) that cite it. Below, every "Anthropic says" is the announcement's own words or a faithful paraphrase, unless noted otherwise.

Its daily rhythm: brief → check → ask you to approve

You don't need a technical background to get Duet — just remember the three-beat rhythm. Each morning it sends you a morning brief, summarizing "what matters most today" across your connected apps. During the day it checks those apps every few hours and comes to you when something needs your eyes. And for anything irreversible — sending a message, paying, signing, deleting — it lays the exact draft in front of you first and waits for your yes. Anthropic's own phrasing: "checks with you before it acts on your behalf," and the showing must reach "shows you exactly what it's about to do."

The most savor-worthy of the three beats is the second one: "checks every few hours," not real-time monitoring. That isn't a technical limitation — Anthropic did the math: checking nonstop "would use up your plan's limits much faster." Duet runs on Claude Opus 5.5 (no model switching during the preview), and every background check burns model turns; those turns all count against your existing Pro/Max usage limits, with big tasks — a financial model, a deck — costing even more. So "a glance every few hours" is the quota-economics optimum: it feels like "it's always watching for you," without torching your quota at 3 a.m. That detail is the most honest cost reference any indie developer building an always-on agent could ask for — we'll expand on it below.

Who can use it now: the three gates of the waitlist

A bucket of cold water first: most readers of this piece can't use Duet yet. The research preview has three gates. Gate one is geography: US only. Gate two is subscription: Claude Pro or Max — the free tier doesn't qualify. Gate three is the invitation: join the waitlist and invitations go out in waves, first-come not necessarily first-served, with each invitation tied to the email on your Claude account. Per third-party summaries of the official terms, Duet is included in Pro/Max plans at no extra charge — but its background checks and big tasks draw from the same usage pool you already have. In other words, it competes with your everyday chats for the same quota; if the quota runs dry at month's end, don't blame it.

Team and Enterprise are not supported yet; Anthropic only says "later." That ordering is itself informative: Anthropic gave Duet to individual subscribers first, not enterprise customers. The reason isn't hard to guess — a proactive agent's "art of interruption" and approval chain need tuning in a small, high-tolerance population first; enterprise compliance, audit, and permission systems aren't ready to receive an agent that "messages people on your behalf at midnight." My judgment: Duet's research preview is, in essence, a large-scale user-feedback harvest — Anthropic itself says it will "keep tuning its judgment based on what testers tell us." So the US users joining the waitlist now aren't getting a finished product; they're getting a ticket to help Anthropic tune the parameters.

The substrate: you shut down, it keeps going

Duet gets its own cloud virtual machine and browser. Close the laptop, pocket the phone — it keeps working in the cloud and notifies you when the task is done or needs your call. That is the fundamental break from "Claude in a chat box": the chat stops when you leave the conversation; Duet is "give it one sentence, it goes and does it, then reports back" — asynchronous work. Anthropic puts it plainly: "Work continues when you step away."

Migration cost is near zero: Duet inherits your memory, skills, and connected apps from the same Claude account, so setup is minimal. It also brings 6,000+ connectors and plugins from the Claude Marketplace, and the scenarios Anthropic names all orbit "the daily chores of a small-business owner": finding Gmail and Slack messages you missed and drafting replies; checking QuickBooks or Stripe for unpaid invoices and drafting reminders; preparing payroll in Gusto with hours, bonuses, and reimbursements filled in; preparing agreements and sending them for signature through DocuSign; writing follow-up emails from HubSpot contact and deal history; turning a product launch into a Canva social campaign; setting up and managing a Shopify store. It can even buy things for you: via Link, Stripe's digital wallet, it confirms the item with you and pays with a virtual card, so your real card number is never exposed; it can check out at online stores with Shop Pay. The next wave is already previewed: with your permission, it will sign in to sites with logins you saved in 1Password.

Permission design: it can spend your money, but it can't trade your stocks

Duet is the most explicit of the three about approvals. The rule is clean: reading, drafting, preparing — it does those itself; sending, paying, signing, deleting — it shows first and waits for your confirmation. But Anthropic drew a red line that holds even with your approval: securities trading and moving money between accounts are off-limits, period. That preemptively locks down the biggest trust landmine of proactive agents — an agent that can message people at 3 a.m. and touch your bank accounts, if it could also trade stocks and shuffle money, would be one prompt injection away from financial catastrophe. The ban isn't timidity; it's honesty.

The supporting safety design has three layers worth reading closely. First, credential isolation: Duet's credential handling follows the Claude Managed Agents design — it never directly sees your passwords or sign-in tokens. It can operate your accounts without holding the keys. Second, action approval runs through the auto mode classifier, the same safeguard that keeps Claude and Claude Code from taking steps you didn't ask for. Third, a sensitive-information classifier: a newly released dedicated classifier plus a safety system prompt checks each message; the strictest tier — credit card numbers, government IDs — is rejected outright for sharing; health or personal-finance information goes through only when you have explicitly instructed Duet to send it. Anthropic adds one honest sentence: "We're continuing to develop and strengthen these protections, and will share more details about how Claude Duet is designed for safety in the coming days" — an admission that the protection is a work in progress, not a finished product.

An interesting contrast: Duet looks powerful, but there is plenty it can't reach. It never holds your login credentials (the keys stay with the managed-agents infrastructure); and during the preview it can't show you its browser — it browses on your behalf, but you can't see the screen (Anthropic admits "it can't yet show you its browser"). That inverts the "seeing is trusting" instinct: its trust anchor isn't transparency, it's the approval chain of "show the exact draft, wait for your yes." Whether that design survives the real world's prompt injections will depend on the coming red-team reports — Anthropic says an external firm has penetration-tested it and Duet is covered by the public bug bounty, but the findings arrive "in the coming days."

The wedge comparison: three proactive agents, three anxieties for sale

"The agent comes to you" — the three vendors say the same sentence, but they sell three different anxieties. Duet's target user is written with unusual specificity: "professionals and small business owners with more on their plates than they can track." The named people in the announcement back it up: Boris Cherny, who leads the Claude Code team, talks through his to-do list with Duet on walks; Chelsea Larsson, who leads model and product behavior design, hands Duet her list before leaving to be with her kids and gets an evening report; Karan Ahuja, who leads the Duet team, uses it to schedule user interviews, prototype, and even ship code. In the small-business early access, one 110-person restaurant (per a third-party retelling of the official early-access stories) uses it for invoicing, customer support, and marketing.

Dimension	OpenAI Dots (covered earlier)	Meta Muse (covered earlier)	Anthropic Duet (this piece)
One-line positioning	OpenAI's proactive-agent play	Meta's proactive-agent play	Anthropic's proactive agent, aimed at "the daily chores of small-business owners"
Underlying model	See the Dots piece	See the Muse piece	Claude Opus 5.5, no switching during the preview
Trust-design keyword	See the Dots piece	See the Muse piece	"Show the draft, wait for your nod" plus a forbidden zone for securities/transfers
Cost narrative	See the Dots piece	See the Muse piece	The quota math of "checking every few hours" written into the official announcement

The subtext of this table: all three have abandoned the old battlefield of "model-parameter racing." Duet even takes the model choice away from you — Opus 5.5 or nothing, not even the chance to compare. That is rare confidence (or arrogance) in AI product history: Anthropic bets that in the proactive-agent category, users pay for "peace of mind," not for "a smarter model."

So Duet's fight with Dots and Muse isn't on technical specs — it's on "whose time it saves." Its product logic is attention scarcity: help you decide what to look at today, and reserve your energy for "the ambitious projects that felt impossible before." The competition in proactive agents is no longer "whose model is smarter" (Duet locks you to Opus 5.5 — it won't even let you compare) but "who understands one kind of person's daily life better": Duet bets on the small-business owner's inbox and ledger. My judgment: this wedge is Anthropic's smartest move and its hardest. Smart, because "watch my QuickBooks for unpaid invoices" is an easier sell to non-technical users than "write my code" — it's a plain trade of money for time. Hard, because small-business owners have the lowest error tolerance: one missed invoice, one misfired dunning email, and real money is lost. That's why Duet's approval chain has to be more conservative than any developer tool — "show the draft and wait for your nod" isn't a nice touch; it's the survival line.

Quota economics: a cost reference for indie developers building always-on agents

Back to that "checks every few hours" detail, because it deserves its own chapter. Anthropic's accounting is transparent: background checks count against your existing Pro/Max usage limits, and bigger tasks burn more model turns. Translated into indie-developer language: the "always" in "always-on agent" is priced per token, per minute. Want it watching ten apps 24/7? Fine — your quota burns first. Anthropic's design of "poll every few hours, disturb only when needed" compresses "always online" into something a subscription can bear.

That is a ready-made pricing lesson for indie developers shipping agent products. Many "AI assistant" demos sell "real-time responsiveness," but the production ledger is brutal: poll interval × tokens per check × number of apps = your monthly hard cost. Duet's answer downgrades "real-time" to "every few hours" and reserves the saved quota for the big tasks that genuinely need model turns. If you're building an always-on agent, Duet's formula is your starting point: first compute "what one check costs," then decide "how often to check," and only then talk about "real-time." Anthropic even wrote the logic into its user-facing copy — "checking nonstop would use up your plan's limits faster" — which is rare in vendor announcements: vendors usually hide cost behind "infinite possibilities" narratives, while Anthropic made cost part of the feature design. Honest — and also a sign that the subscription quota ceiling is the real bottleneck of the proactive-agent category.

Limitations Anthropic admits: read the "can't" list before the "can" list

The announcement carries an "this is still an early version" section worth reading line by line, because it draws Duet's capability boundary:

• It may misread urgency and misread threads. It "might misread a thread or miscalibrate what's urgent." The biggest risk of a proactive agent isn't "doing nothing" — it's "bothering you at the wrong time with the wrong priority." An agent that wakes you at 3 a.m. over an invoice reminder is more annoying than one that does nothing.
• It misses things between checks. The flip side of "checks every few hours": anything new that happens between checks, it doesn't see. Scenarios that demand real-time response (customer support, trading) are out of reach for now.
• You can't see its browser. It browses on your behalf but can't show you the screen during the preview — you only know "what it's doing" through its reports. A big hole in transparency, and the update most worth watching in later versions.
• No phone or SMS; Microsoft 365 not available yet. The office stack enterprise users live in doesn't connect right now; personal errands like booking tickets are, by Anthropic's own admission, "not quite as strong."
• Quota is a hard constraint. Background checks and big tasks both eat your Pro/Max quota — heavy users may find "the harder it works, the faster my quota burns."

What makes this "can't" list precious is that the vendor wrote it itself. Announcements usually only list the "cans"; Anthropic put "misreads urgency," "misses things," and "can't show its browser" into the launch post — a preemptive inoculation for users, and an "honesty" benchmark for competitors. My judgment: the real exam of Duet's research preview isn't feature count but "the art of interruption" — it must find each user's sweet spot between "missing the important thing" and "over-interrupting," and that sweet spot differs for everyone. Anthropic says "we'll keep tuning its judgment based on what testers tell us" — translated: the core algorithm of this product isn't the model; it's the user-feedback data on "when to stay quiet."

Three judgments for vibe coders

First, proactive-agent competition has moved from "models" to "scenarios + trust design." Duet locks you to Opus 5.5 and won't let you pick a model — a declaration that in this category, the model is plumbing; differentiation lives in "whose time, and which kind of time, you save" and "how trustworthy your approval chain is." Indie developers building agent products: stop racing on "plugged in a bigger model" and race on "what keeps my user up at night."

Second, "show the draft, wait for the nod" is table stakes; the forbidden-capabilities list is the moat. Duet's "no securities trading or transfers even with approval" belongs in your product requirements doc: some actions are risky not because "the user didn't confirm," but because "the confirmation itself may have happened under manipulation." Giving the agent a capability forbidden zone matters more than giving the user a confirm button.

Third, the subscription quota is the proactive agent's ceiling — and your pricing reference. Duet counts background checks against Pro/Max quotas, making "always-on" costs transparent. If you're building a consumer always-on agent, Duet's "poll every few hours" is your cost-floor reference; if you're building for business, when a client asks "how much does it cost to watch around the clock," you can lay out Duet's quota ledger: poll interval, tokens per check, number of apps — multiply the three and there's your answer.

One last line: Duet is only a research preview right now — US Pro/Max subscribers on a waitlist, no Team or Enterprise in sight, and users outside the US can only watch. But "the agent comes to you" has gone from Dots to Muse to Duet — three vendors turning a concept into products. Once a paradigm is established, what remains is the long race over "who understands your daily life better" and "whose approval chain you trust more." Duet's opening bet is the small-business owner's ledger and inbox — whether it wins depends on how fast it learns "when to stay quiet." And for those of us building agents, the most valuable thing isn't Duet itself — it's the ledger it published in the open: proactivity is billed by the token.