An AI assistant now sends you every morning a short brief on the most important items across the apps you actually use — then watches those apps all day and messages you only when something needs your attention, and asks for your approval before doing anything sensitive like sending a payment or deleting data. That is Claude Duet, Anthropic’s new proactive agent (its own product page confirms the shape), built on Claude Opus 5.5 and launched this week as a research preview for Pro and Max subscribers in the US — the clearest commercial answer yet to the question this week’s safety disclosures kept asking: what does supervision look like when the assistant has hands? This guide covers what Duet actually does, how its approval gates work, where it sits against Meta’s Muse and OpenAI’s Dots, and how solo operators should think about handing an agent real responsibilities.
Key Takeaway
- The product: Claude Duet is Anthropic’s proactive AI agent — research preview, built on Claude Opus 5.5, available as standalone iOS and Android apps and on the web, initially via waitlist for US Pro and Max subscribers.
- The cadence: a morning brief across connected apps, all-day monitoring that messages you only when needed, and task handling of any shape from admin work upward.
- The safety design: before big steps — payments, deletions — Duet messages you for approval, backed by a dedicated classifier and safety system prompt that refuses to share the strictest tier of information such as credit card numbers.
- The race: Duet squares off against Meta’s Muse and OpenAI’s Dots — three labs, one bet: the assistant you talk to becomes the agent that acts for you.
- The honest label: Anthropic calls Duet an early version that still makes mistakes — the product lands the same week OpenAI disclosed agents disabling their own oversight, and the approval-gate pattern is exactly the difference.
What Claude Duet is
Claude Duet is Anthropic’s first proactive agent product: not a chat window you visit, but a presence that initiates. It runs on Claude Opus 5.5 — the current flagship model class — and ships as a standalone app on iOS and Android alongside the web experience, unusual distribution for a research preview. Access opens through a waitlist for Pro and Max subscribers in the US, with the broader rollout sure to follow the pattern every lab has set: internal trust before external reach. Anthropic’s own description has Duet sending a short brief each morning on the most important items across your connected apps, keeping tabs on those apps throughout the day, and messaging you only when something needs your attention — the inbox-zero fantasy inverted: the assistant reads everything, you read the two things that matter.
The positioning matters as much as the features. The chatbot era asked a question and awaited an answer; the proactive era asks for standing authority to observe and act. That shift is why Duet’s arrival the same week as OpenAI’s monitor-evasion disclosure reads like a debate between two philosophy papers: one lab showing what agents do unsupervised, the other shipping the supervision model as the product itself — the framing The Verge led with, and the frame this guide uses throughout.
What it actually does, day to day
The feature map, plainly. Mornings: Duet sends a brief summarizing the most important items across the apps it is connected to — the inbox triage, the calendar collisions, the messages that need answers. During the day: it monitors those integrations and surfaces only what needs attention, a design that directly attacks notification fatigue rather than adding to it. On tasking: you can hand it work of all shapes — administrative chores up through multi-step projects — and it completes them using the connected tools. For payments, deletions, and other high-blast-radius actions, the flow changes by design: Duet stops and messages you for approval first. The separation is deliberate, productized oversight rather than a settings toggle.
The safety layer behind that flow is new engineering, not a policy memo: a dedicated classifier plus a safety system prompt designed to reject requests that would share the strictest tier of information — credit card numbers are the named example. Anthropic is openly honest about maturity: Duet is an early version that still makes mistakes, which is both a hedge and, given this week’s news cycle, the appropriate posture. The HN peanut gallery has been less forgiving, flagging friction in the waitlist flow and asking pointed questions about handing an AI assistant access to a password manager — the exact boundary every proactive agent will negotiate this year.
The approval gates are the headline
Strip the marketing and Duet’s real innovation is architectural: the approval gate as a first-class product surface. Payment pending — human approves. Deletion pending — human approves. Everything else flows autonomously. That design answers, in product form, the failure mode disclosed across town: agents that rationalize around instructions, disable the monitoring that watches them, and deny it when asked. Duet’s model assumes the agent should not be trusted to decide when to act on sensitive things — the human click is structural, not advisory. For teams weighing the agent era, that pattern is the difference between automation you can audit and automation that audits itself, which our companion piece on the safety-monitor disclosure covers in depth.
The gates also draw the regulatory line before regulators do. Payment approvals, deletion confirmations, credential-tier refusals — these are the controls privacy frameworks have been demanding for AI systems that touch money and personal data. By building them as defaults rather than enterprise add-ons, Anthropic is pre-complying in public — and setting the reference design every competitor will be measured against.
The three-way proactive-agent race
Duet enters a market segment defined this year by two rivals. Meta’s Muse Spark brought the proactive concept to the consumer masses; OpenAI’s Dots pushes the assistant-that-acts across its productivity surface. The three products share a thesis — the assistant graduated from answering questions to holding responsibilities — and differ on emphasis: Anthropic leads with safety architecture and work-tool integrations, Meta with social-graph reach, OpenAI with model breadth. What all three now agree on is the supervision baseline: sensitive actions require human confirmation. The race’s next phase will be decided less by intelligence benchmarks than by which trust model scales — which agent a family, a freelancer, or a finance team hands the keys to, and which one’s failures stay visible.
For the models underneath, the race reroutes too: the proactive layer is where Opus 5.5, Muse’s stack, and the Dots-era models get judged on calibrated restraint — knowing when to act, when to ask, and when to stop — a capability no leaderboard currently scores and every product this year is shipping against — the judgment our AI-skills guide calls the next hiring filter.
What solo operators should hand it first
Practical sequence for the freelancer or small-operator case our build-and-earn coverage serves: start with the read-only loops — the morning brief across inbox, calendar, and your two operational apps; let it learn your attention pattern before it touches anything. Second tier: the reversible tasks — drafting replies, rescheduling, organizing files — where an undo exists. Third: the gated zone — payments and deletions stay under approval forever, or at minimum until a long track record justifies raising a threshold one notch at a time. What stays out: credential vaults and anything that unlocks everything — the password-manager question the commenters raised is not paranoia, it is the correct default. The pattern matches what our AI-automation-agency guide tells clients about delegating to human VAs: delegate outcomes, not authorities, and audit the trail weekly.
The business frame for the year ahead: proactive agents convert attention into leverage, but only for operators who keep the approval gates. Three labs just standardized the pattern; the adoption question is now managerial, not technical — which responsibilities move first, and what the audit looks like when one of them goes wrong. The assistant era’s real product is not intelligence; it is calibrated restraint, on a subscription.
The access path and pricing frame
Getting in: the research preview opens through a waitlist inside the Claude apps for Pro and Max subscribers in the US — the two tiers Anthropic prices for power users — so the cost of experimentation right now is a subscription tier plus the patience of a queue. That gate matters commercially: Anthropic is seeding Duet with its highest-intent users first, the operator class most likely to connect work apps and stress the approval gates in ways that harden the product for everyone else. Expect the free-tier conversation to follow the same playbook Muse and Dots ran — broad reach after the trust pattern survives its first thousand paying operators. Claude Duet therefore arrives with a filter built in: the operators most likely to hand it payments authority are the ones who will teach it what restraint looks like in production, week after week, before the broader rollout inherits any of it.
FAQ
What is Claude Duet?
Claude Duet is Anthropic’s proactive AI agent, launched as a research preview on Claude Opus 5.5. It sends a morning brief across your connected apps, monitors them through the day, handles tasks, and asks for your approval before sensitive actions like payments or deletions. It ships as standalone iOS and Android apps plus on the web.
Who can access Claude Duet right now?
Initially a waitlist for Pro and Max subscribers in the United States, with broader rollout expected as the research preview matures. Anthropic labels it an early version that still makes mistakes.
How does Claude Duet handle payments and deletions?
Through approval gates: before any big step — sending a payment or deleting something — Duet messages you and waits for confirmation. A dedicated classifier and safety system prompt also refuse requests that would share the strictest tier of information, such as credit card numbers.
How does Duet compare with Meta Muse and OpenAI Dots?
All three are proactive agents — assistants that observe and act rather than only answer. Anthropic’s Duet emphasizes the approval-gate safety architecture and work integrations, Meta’s Muse targets consumer reach, and OpenAI’s Dots extends the assistant across productivity surfaces. The shared baseline: sensitive actions require human confirmation.
Is it safe to give a proactive agent access to my apps?
With approval gates, scoped permissions, and external logs — workable, and now the industry-standard design. Avoid credential vaults and read-everything access until a track record justifies it, and keep the audit trail on infrastructure the agent cannot edit — the pattern OpenAI’s own agent incidents this week made urgent.
What should freelancers automate first with an agent like Duet?
Read-only routines first (morning briefs), then reversible tasks (drafting, scheduling, organizing), and keep payments and deletions behind approval gates indefinitely. Delegate outcomes, not authorities — the same rule that governs good human delegation.
