Claude Duet: Anthropic's Proactive Agent, Powered by Opus 5.5

Anthropic launched Claude Duet, a proactive personal agent that works on its own cloud computer, briefs the user each morning, and pauses for approval before sensitive actions. It is powered by Opus 5.5, ships as a standalone iOS and Android app plus the web at claude.ai, and is gated behind a waitlist for Claude Pro and Max subscribers in the United States as a research preview.

What Claude Duet is

Claude Duet is an agent you hand work to by message, which then runs in the cloud rather than in a chat window you have to open. Anthropic's product page frames the pitch in one line: "Get back to what needs you." The agent takes on tasks of all sizes, keeps working while the user is away, and waits for confirmation when an action matters.

Three product promises define the offering:

• It knows what needs attention, sends a morning brief, drafts next steps, then goes quiet until something requires the user.
• It takes on large tasks from a single message, such as setting up a meeting or building a dashboard, and reports back when finished.
• It carries over the user's existing Claude memory, skills, and connected apps from the first message, inside one thread.

Key facts at a glance

Attribute	Detail
Product	Claude Duet
Vendor	Anthropic
Status	Research preview, waitlist
Underlying model	Opus 5.5
Platforms	Standalone iOS and Android app; web at claude.ai
Plans with access	Claude Pro and Max; Team and Enterprise not included yet
Region	United States only
Integrations	More than 6,000 connectors and plugins
Input modes	Text or voice
Execution environment	Its own computer in the cloud

Anthropic's connector gallery for the product shows Stripe, 1Password, Shopify, Gmail, Google Calendar, Slack, Intuit QuickBooks, Docusign, Google Drive, Hubspot, Gusto, Microsoft/Office 365, Notion, Canva, and Ramp.

How Claude Duet differs from the Claude app

Anthropic's own FAQ draws the distinction directly: the Claude app waits for the user to open a chat, and the user works alongside Claude. Claude Duet comes to the user instead. It surfaces what needs attention each morning, accepts handed-off tasks by message, and reaches out when a task is complete or requires approval.

Claude Duet is not a separate account or a separate memory store. It uses the same Claude account, including memory, skills, and connectors, and keeps history in a single thread.

The Marlo demo: what the product looks like in practice

Anthropic's walkthrough follows a fictional coffee company founder, Maya, running a chain called Marlo. The demo shows the full loop rather than a single task:

• A morning brief covering hiring, a lease negotiation, a location opening, a fundraising round, and delivery errors.
• A 41-item Chicago opening checklist built from Maya's two previous openings, with 33 items done or on track, 7 being watched, and 1 requiring her approval today: a $640 county health permit fee.
• Series A preparation for Harbor Ventures, including a 12-slide deck rebuilt with Q3 data, a data room at 23 of 26 documents, and background research on the six Harbor partners.
• A standing dashboard, Marlo Pulse, that pulls from point of sale, Shopify, CRM, support inbox, payroll, and bank data, and refreshes at 5:30 AM.
• A proactive finding: East Bay churn rose from 2.6% to 3.4% after a shipping price change, with a proposed two-week test of free shipping over $40.

In the demo, every consequential step ends at a human decision. The agent drafts an email to a Harbor partner and asks "Send from you?" before delivering it.

Approval and privacy model

Anthropic states that Claude Duet asks before sensitive operations such as sending messages or paying, and that approvals, logins, and safety checks run in the user's own cloud computer so sensitive details stay private. The product page also emphasizes visibility: the user can see what Claude Duet has done and is doing, and steer it.

That design claim is the sharpest point of contention in the Hacker News discussion, covered in detail below. One commenter summarized the credential-sensitivity concern this way:

This thing apparently has an integration with 1Password. Am I hopelessly out of touch in this brave new AI world, or is giving an LLM access to your password manager the mother of all bad ideas?

Availability, pricing, and usage limits

Claude Duet is a research preview on a waitlist for Claude Pro and Max subscribers in the US. Team and Enterprise plans are not included yet. The waitlist form asks for first name, last name, email, company name, phone platform (iPhone or Android), and what the user would hand off to Claude Duet first.

On usage: during the research preview, Claude Duet shares the plan's usage limits with the rest of Claude. Anthropic notes that Duet consumes those limits differently than a typical conversation, and that its background checks count toward them. Larger tasks such as building a financial model or a deck require more model turns and background work.

Known limitations Anthropic acknowledges

Anthropic lists three limitations plainly on the product page. Claude Duet may misread a thread. It may miscalibrate what is urgent. It can miss something that arrives between checking in with the user.

How Hacker News reacted

The submission collected 485 points and 446 comments on Hacker News. The reaction split along four lines: whether proactive agents have a real moat, whether users should hand a model access to their accounts, whether a closed hosted agent is the right architecture at all, and whether the industry is shipping the same product four times over.

Positioning: a crowded personal-agent field

Commenters immediately placed Claude Duet alongside OpenAI's Dot, Meta's Muse, and other agent products, with several calling it a direct competitor to OpenAI Dot and to self-hosted agent stacks such as OpenClaw. One widely echoed summary: "It's an OpenAI Dot/Open Claw competitor. Doesn't really seem to differentiate itself meaningfully at first glance."

A separate thread noted the deliberate branding difference. As one commenter wrote:

A notable distinction between Duet and Meta Muse/OpenAI Dots is the lack of quirky anthropomorphic branding. As a result, the use cases are more clear/more professionally-oriented as they don't rely on you creating a magic pet/BFF that does whatever you want.

The timing drew its own commentary, with one commenter observing that "dot, bot, and duet, all announced in a month. you couldn't make up this level of industry lockstep. now all we need is google bit and microsoft ditty."

The margin argument: proactive agents as batch-filler

One of the most technically interesting comments argued the economics matter more than the product surface. Because a personal assistant does not have to answer in real time, its work can be scheduled as low-priority batch jobs:

I suspect that Dots and Duet are very high margin drivers for the labs. Personal assistant chat sessions don't have to be realtime, so workloads can be scheduled on low priority batches and be waited if higher value work comes into a queue. Inference providers are always trying to fill inference batches and having a stream of low priority requests can help pack batches that might have gone empty, without raising the latency of higher priority requests such as live chat sessions.

Trust and credentials: the 1Password question

Several commenters said they would not use an always-on agent because it cannot reliably infer which actions they would approve. One framed the issue as a hard blocker rather than a calibration problem:

Love LLMs, use them extensively, but they cannot intuit what actions they can take on my behalf that I will or will not approve of, and therefore I will never use one of these things.

Another questioned the premise of the whole category: "Are people really okay giving these agents access to all areas of their lives? And what do you even use it for?" A third raised the compounding-data risk of delegating broadly:

As much as I would love to make the most of these tools, I am not confident that yielding control of my personal data won't hurt me - and this seems like table-stakes, based on the type of value everyone is getting from them.

Lock-in, open source, and "don't subcontract your life"

The most forceful architectural counterargument was that the value of an agent comes from the infrastructure underneath it, not the chat surface, and that the infrastructure should not belong to a single vendor:

Muse (and Claude Duet, and Dots) are slippery slopes from which it will be very hard to climb back. Do not subcontract the managing of your life to Facebook, Anthropic or OpenAI (or to anyone else for that matter). Intelligence has to become a commodity, the infrastructure via which the agents do useful work has to be open source, and the data and learnings and SOPs they produce have to be yours.

Others already run locally-controlled equivalents. One commenter described running a proto version of the product using Claude Code and cron jobs for an online jewelry business, and specifically valued that Claude Code uses locally stored memories because it keeps the option of switching agents open. Another runs a Hermes agent locally with a mix of local models and inexpensive hosted models, and described a likely market for consumer agents that live on a home network rather than a vendor cloud.

Lock-in outside the model layer appeared too. One commenter said that if they want a persistent agent, they are still unsure why they would choose one tied to a specific model family: "Having it 'just work' is nice! But this feels like a product area where not being locked into one garden is more important and the net gains are greater."

Interface matters more than the model

A recurring theme was that the winning interface for a personal agent is the messaging app people already use. One commenter described a family deployment in detail:

I wish it interfaced via iMessage. I don't want to train the family on yet another app. We run a Hermes agent on an old MacBook Air. We invited the agent "Rosey" to our family chat... She does our meal planning and grocery shopping at Walmart via BrowserMCP... It adds/updates calendar events via icalbuddy cli. Reads balances from credit cards for budgeting... iMessage has been the perfect interface fit for us.

That tradeoff is explicit in Claude Duet's design: it is a new standalone app, not a presence inside an existing messaging client.

Browser blocking, CAPTCHAs, and the honesty problem

Multiple commenters said web automation is the weak point of any hosted agent, because sites block, rate-limit, or CAPTCHA automated traffic. One described the downstream failure mode:

I don't trust Claude as personal assistant yet because too many websites block it from reading their contents, leaving Claude to guess and rely on very indirect sources. Performing actions online is even worse, with CAPTCHAs and whatnot. It has made enough mistakes in this area that I now have a standing instruction (the only one) to always loudly warn whenever the official source for something couldn't be fetched.

A proposed fix, from another commenter, was to separate the agent from the browser entirely: an independent personal data hub and a browser MCP that the assistant uses to act as the user, giving the user visibility and control over when cookies or credentials are used, and federating the traffic so blocking is less likely.

One commenter running a local agent on a 128GB RTX Spark for a family of four argued the throughput requirements are lower than people assume, and that operating at human-like speed avoids the automation signals that standard webdriver-based tooling emits.

Interface and product-design critiques

Some of the sharpest criticism concerned the waitlist itself rather than the agent. One commenter tried to sign up from the Android Claude app and reported that Claude said it could not do so:

I tried asking Claude fron the android app if it could sign me up for the waitlist and it said that it couldn't. There's no technical reason this shouldn't be possible, it just shows a lack of care and attention to detail.

A related structural critique was that Duet should not be a separate app at all: "These AI companies appear to be 'shipping their org chart'. Shouldn't this just be native to Claude, not a separate app/UX...?" Another commenter argued the demo persona would already have an executive assistant or staff doing the work shown, which makes the value proposition harder to read for ordinary users. Several people simply asked for a concrete explanation of what Duet does that a normal Claude or ChatGPT session does not.

Fatigue, sameness, and the "org chart" critique

A durable undercurrent in the thread was weariness with AI announcements dominating the front page, and with the sense that every frontier lab converges on the same product:

We have this super power technology and somehow every single AI company end up building the exact same product... I guess that's what happens when you delegate your decision making and development process to LLMs at scale, you end up with soulless products that all look pretty much the same and lack actual human ingenuity.

Others raised narrower concerns: the US-only launch drew a pointed complaint, one commenter noted that Claude Duet will be pointless "unless the overzealous guardrails are fixed," and another argued the product copy overpromises relative to existing customer service and citation quality.

Enthusiastic counterpoint

Not everyone was skeptical. One commenter described having a frontier-model assistant handle the purchase of a car and, at the time of writing, a house purchase in Kyoto in a language they do not speak, coordinating agents, scriveners, a hanko stamp, inspections, and insurance:

I find having a personal assistant that's using a true frontier model makes a huge difference in how much I can just trust the output without doing the work again myself just to double check the results.

Another, comparing OpenAI's Dot to Anthropic's models, said Dot was worse than expected at anything coding-related but genuinely useful as a monitor of communication channels, and that they would switch to Duet partly because they prefer Anthropic's models and already hold a larger subscription there. A third commenter raised a labor-market angle, noting that much of the current advantage of skilled users comes from capabilities available only through Claude Code and desktop clients, and that broad availability through a consumer agent changes that dynamic.

What to watch

Three open questions will determine whether Claude Duet is a durable product or another entry in a converging category.

Whether proactive agents have a defensible moat. Commenters largely agreed the product surface is easy to copy; the disagreement was whether the moat is inference economics, memory and integration depth, or nothing at all.

Whether approval UX is sufficient for credential-sensitive integrations. Anthropic's stated model is human approval at sensitive steps, but several users said the harder problem is that an agent cannot predict which actions a person would refuse.

Whether hosted agents can operate on the open web. Blocking, CAPTCHAs, and unreliable source retrieval were the most frequently cited practical failure modes, and the proposed alternatives in the thread all involved moving the agent away from the vendor's browser entirely.