Claude Duet: What to Delegate, What to Keep

The expensive part of running a business is often not writing the reply. It is remembering that the reply is needed, finding the promise in last week’s thread, checking whether the invoice has been paid, and deciding whether today is the right day to follow up. A chat assistant helps once you assemble that context. A proactive agent promises to assemble it before you ask.

On October 21, 2026, Anthropic introduced Claude Duet: a personal agent that works in connected apps, keeps going on its own cloud computer, and returns when work is ready or a decision needs you. Here is what the official research preview actually offers, how to start without over-delegating, and where the appealing story needs a hard boundary.

What Claude Duet is—and what it is not

According to Anthropic’s getting-started guide, you message Duet through text or voice. It uses your Claude account’s memory, skills and connectors, works on a computer in the cloud, and can continue after you close the app. A morning brief draws on email and calendar; larger assignments can produce a research brief, financial model, deck or dashboard.

The interaction has three useful surfaces: Chat for conversation, Tasks for work in progress and decisions that need you, and Make for finished files. Separate topic channels share memory. Duet can suggest that a task is finished, but only you can mark it completed. That last distinction is good product design: an agent’s confidence is not the customer’s acceptance.

This is not Claude Code, the development tool, and it is not the unrelated open-source project with the same “claude-duet” name. This article covers Anthropic’s October 2026 product. There is no terminal installation command to run here. The official product page describes standalone iOS and Android apps and web access at claude.ai.

Why the shift matters: from a prompt to a pending decision

A conventional assistant waits for the next prompt. Duet’s proposition is that preparation can continue between prompts: detect a follow-up, draft it, link the supporting records, then ask for a decision. The potential benefit is fewer context-reconstruction sessions, not simply faster text generation.

For a solo operator, that could mean arriving at a client meeting with the open questions already assembled. For a product leader, it could mean a weekly brief that exposes missing evidence rather than another attractive dashboard. In enterprise product work, the recurring problem is not a shortage of charts; it is disagreement about what the chart represents, which promise is binding, and who owns the next decision.

My product judgment is that Duet becomes valuable when it reduces those coordination costs without hiding the underlying state. It becomes a distraction when every background observation generates another notification. The right denominator remains accepted work: useful output after review, corrections and consequences—not the number of tasks an agent claims to have handled.

Claude Duet vs other assistants: compare the job, not the model

“Which model is better?” is the wrong first question here. Claude Duet, OpenAI Dots and Meta Muse are assistant products: memory, connected apps, background work, controls and a user experience wrapped around an underlying model. A more capable model does not automatically give you the right approval boundary, business access or dependable notification behavior. We have not run a head-to-head performance test, so this is a comparison of documented working relationships—not a leaderboard.

Evaluate Gemini Tasks for Google-centered ongoing work. Google’s current task guide describes connected sources, scheduled workflows and browser-assisted tasks. It requires a personal adult account, Google AI Pro or Ultra, Keep Activity and Personal Intelligence; work and school accounts are excluded. A remote browser can continue after your device closes, while local Chrome work requires the device awake. Google warns against sensitive offline schedules; do not assume every consequential action will pause in precisely the same way as Duet.

Choose Duet for a narrow Claude-centered business pilot. Its morning-brief and approval-oriented design fits the recurring preparation workflows above. The trade-off is a US waitlist, shared subscription capacity and no browser takeover in this preview. If your company needs Team or Enterprise access, Duet is not currently that option.

Evaluate OpenAI Dots when you want to inspect the computer or configure action rules. OpenAI’s official guide describes an always-on agent in ChatGPT with scheduled work, visible computer interaction, optional local access and four custom-rule behaviors. That flexibility is not automatically safer: a no-ask rule expands authority, and local-computer access expands the data surface. Keep local access off unless the task genuinely requires it.

Evaluate Meta Muse for a lower-cost personal experiment. Meta’s onboarding guide documents free usage with limits, directed tasks, reminders, shared-context side chats and a Library for artifacts. It also explicitly warns about unexpected actions. The guide does not establish the same sensitive-action approval rules as Duet, so do not assume parity merely because both are called personal agents.

Keep ordinary Claude, ChatGPT or Gemini conversations for bounded thinking. If you mainly want to critique a strategy, summarize an uploaded document or refine a draft while you are present, you may not need an always-on agent at all. Connected-app and scheduled capabilities vary by product, plan and region; a conversational assistant is not proof of Duet-like background delegation. For software changes, Claude Code is a development workflow, not an interchangeable personal assistant.

One current naming trap: OpenAI’s legacy ChatGPT agent help page now says agent mode is no longer available and points to ChatGPT Work and cloud-browser workflows, while older instructions remain farther down the same page. Do not use those older quotas or the /agent command as today’s buying guidance. Check the current product your account actually offers.

The decision rule: choose the smallest authority and cheapest access that produces a verifiable result in the apps you actually use. Compare on the same brief, inspect source accuracy and record review effort. Treat vendor safeguards as claims to test—not a reason to switch off judgment.

Access, cost and the limits worth reading first

• Eligibility: US Claude accounts on Pro or Max, with a waitlist invitation. Team and Enterprise are not included yet. Joining does not guarantee an invitation date.
• Preview price: included in Pro and Max at no extra charge during the research preview. This is not a promise of future pricing or unlimited work.
• Shared capacity: five-hour session and weekly plan limits apply. Morning briefs and background check-ins count, so Duet can leave less capacity for other Claude work.
• Model: the guide specifies Claude Opus 5.5. You cannot choose the model; Duet chooses how much effort a task needs.
• Timing: account checks happen at intervals, not continuously. An urgent message can arrive between checks.

Anthropic also documents a one-time full session and weekly limit reset on admission, available until November 30, 2026. Treat that as a dated preview provision, not a recurring allowance. Before changing your subscription, read the current pricing and usage section; “included” describes billing, not the opportunity cost of consuming capacity you need elsewhere.

How to get started without connecting your entire life

1. Join through the official product page. Supply your name, Claude account email, plan and phone type. Access opens in waves.
2. Follow the invitation on your phone. Anthropic says to use the Apple or Google account linked to the waitlist email, then sign in with your Claude account. App availability can take a few minutes after accepting.
3. Review existing connections and memory. Duet inherits them. A new app does not necessarily mean a blank context or a new permission boundary.
4. Start with the minimum useful sources. Email and calendar are a sensible first pilot if you are authorized to connect them. Do not connect business records merely because a connector exists.
5. Hand off one bounded preparation task. Ask for a file or draft, source links, unknowns and explicit exclusions. Inspect it before extending scope.
6. Review Tasks and Make. Confirm the artifact exists and matches the request, then mark completion yourself. A conversational summary is not the deliverable.

The guide lists Gmail, Google Calendar, Drive, Slack, QuickBooks, HubSpot, Docusign and Calendly among supported connections, with limits for some connectors. It explicitly says Microsoft 365 is not available yet—even though the product page displays its logo. For actual preview support, follow the help article rather than treating a logo wall as a compatibility matrix.

Three useful workflows, with a definition of done

Meeting preparation: request a one-page brief separating confirmed commitments, unresolved questions and possible next steps. Require the source thread for each material claim. Done means you can trace the deadline and understand the disagreement, not that the prose sounds polished.

Invoice follow-up: the documented Stripe workflow checks unpaid invoices and drafts reminders for approval. Have it show amount, currency, due date and payment status. Exclude disputed invoices from routine reminders. Done means a correct draft against a current record; sending it is a separate decision. Stripe account access is distinct from Link, its purchase wallet.

Operating brief: have it collect a small set of agreed business measures and flag missing or conflicting records. Ask for observation, source date and uncertainty separately. A dashboard assembled from incomplete data can be more persuasive than a spreadsheet precisely because its design hides the gaps. A causal explanation for falling revenue should be labeled a hypothesis until tested.

These workflows widen what one person can prepare, but they do not remove product judgment. You still choose what matters, which question needs real customer evidence, and which decision is too consequential to delegate. Start with work you already understand well enough to detect a plausible mistake.

Do: delegate preparation with explicit evidence

• Specify the outcome and source scope. “Summarize these connected records for this meeting” is better than “run my business.”
• Ask for contradictions and unknowns. Missing data should be visible, not filled with a confident guess.
• Define escalation. Say which deadline, amount or disagreement needs your attention. Avoid constant updates for routine progress.
• Inspect actual approvals. Verify recipient, attachment, amount, currency, merchant and signature document as applicable—not only the preceding chat.
• Keep a manual fallback. Scheduled checks and usage limits make Duet unsuitable as your only alarm for deadline-critical events.
• Measure your pilot. Track manual time avoided, review time, corrections, missed events and unnecessary interruptions.

Do not: mistake permission for reliable judgment

Do not approve a batch of outward actions because the first draft looked right. A reminder sent to the wrong customer, a misread renewal term or an incorrect purchase total is still your problem. The human sign-off stage can become the new bottleneck, as explored in More Agents. One Person Who Signs Off. Removing preparation work only helps if review stays meaningful.

Do not use Duet as an emergency monitoring service, assume it sees all local files, or expect to supervise its browser visually. The preview guide says you cannot watch or take over its cloud browser; it cannot see all your projects or computer files. It can read work in Claude but cannot change it. It does not call people or send text messages to them, though it can send approved emails from connected accounts.

Do not ask it to trade securities, move money between accounts or access another person’s account. Anthropic lists those as off-limits even with approval. An approved purchase through Link or Shop Pay is a different, narrowly defined workflow—not general banking authority. For a purchase, Duet is designed to ask each time and use a one-time card restricted to that purchase.

Privacy and prompt injection: the quiet risks

A proactive assistant reads content written by people other than you. An email or website can contain a malicious instruction disguised as business information. That is prompt injection: content trying to become authority. Anthropic says Duet is built to treat what it reads as information, but no defense is perfect. Source links and approval gates reduce risk; they do not eliminate it.

Never paste passwords into the chat. The documented browser login flow uses a separate form on your device, designed to send credentials directly to the browser without exposing them in the conversation. Use the legitimate login flow and only accounts you are authorized to access.

Disconnecting a connector stops future access but does not erase memory of earlier conversations. Duet can search its own chats and Claude chats outside projects. Deleting a chat removes it from search; disabling “Search and reference chats” in Settings → Memory stops chat search. Neither statement is a blanket guarantee that every derived record or retained copy is deleted. Our data-retention field guide explains why access, memory, search visibility and deletion must be evaluated separately.

A Pro or Max invitation is also not an organizational authorization to connect employee, customer or regulated data. Check internal policy, contractual duties and applicable terms first. The absence of Team and Enterprise access is a practical reason to keep the initial business pilot narrow, not a challenge to work around.

A seven-day pilot that can actually fail

Choose one repeatable workflow and record a manual baseline. For the first few days, use preparation-only briefs and drafts. Log factual corrections and whether a missing source prevented a decision. Add one low-risk approved action only after the preparation is consistently useful. Keep purchases, signatures and deletion out of the first pilot unless they are the specific workflow being evaluated with appropriate safeguards.

At the end of the week, ask whether you recovered useful human time, trusted the sources, and noticed fewer—not more—avoidable interruptions. If review took longer than doing the work yourself, narrow the task. If an authority boundary failed, pause the pilot and investigate rather than offsetting that failure with an average success score. That is the same release discipline described in agent evaluation before shipping.

The best personal agent does not make you less accountable. It makes the next accountable decision easier to see.

Sources and limits

Researched October 22, 2026, immediately after the launch. This is a documentation-based guide, not a hands-on Duet review: we have not tested an invited account. Capabilities, availability and safeguards above are Anthropic’s documented claims; workflow recommendations and the time model are our editorial judgment. Research-preview details may change.

1. Anthropic: Claude Duet product page (checked October 22, 2026)
2. Anthropic: Get started with Claude Duet (launch October 21, 2026)
3. Anthropic release notes: Claude Duet research preview, October 21, 2026
4. OpenAI: Getting started with your dot (checked October 22, 2026)
5. Meta: How to get started with Muse (checked October 22, 2026)
6. OpenAI: legacy ChatGPT agent page and current retirement notice
7. Google: Manage tasks and workflows in Gemini Apps (checked October 22, 2026)