Most AI assistants wait for a prompt. Claude Duet is designed to reverse that sequence: it checks connected work apps, sends a morning brief, continues assigned tasks in the background, and asks when a decision requires you. The appealing part is not merely that it can draft an email. The difficult part is knowing which message deserves interruption, which task can wait, and when an agent must stop before an external action.
This guide separates the announced product behavior from a proposed trial method. Availability and product claims reflect Anthropic’s public information on October 22, 2026. The scorecard and seven-day plan below are editorial templates, not independent benchmark results.
What has actually launched?
Claude Duet is a research preview with a waitlist, initially for US Pro and Max subscribers. Team and Enterprise are not included yet. It is available as a separate iOS and Android app and through claude.ai. It carries over memory, skills, and connected apps from the Claude account and shares the same plan usage limits. Check the product page and your account rather than assuming the demo is available in every country or plan.
According to the launch announcement, Duet has its own cloud virtual machine and browser. A task can keep running while your laptop or phone is off. It checks connected accounts every few hours, not continuously, and it may misread a thread, misjudge urgency, or miss something between checks. Those are material limits if your workflow depends on real-time escalation.
| Announced behavior | Reasonable expectation | Unsafe assumption |
|---|---|---|
| Morning brief | Highlights from connected apps | Complete, real-time coverage of every source |
| Background work | Drafts and advances assigned tasks | Every complex process completes unattended |
| Proactive messages | Reaches out when a check finds an important item | Zero false positives or missed alerts |
| Approval steps | Displays sensitive actions before they happen | One broad approval makes future actions unlimited |
The public demonstration below places a task queue beside material awaiting review. Keep “a draft is ready” separate from “a message has been sent”; a completed-looking interface alone does not establish the latter.
Public product demonstration, not an independently tested screenshot. The preview interface may change.
Evaluate four stages, not the “AI assistant” label
Use four observable stages: discover, judge, prepare, act. Discovery finds a calendar clash, invoice, or customer message. Judgment decides whether you need to see it now or in tomorrow’s brief. Preparation gathers context and drafts a response. Acting changes the outside world. A trial can create value in the first three stages even if you never grant broad execution rights; the fourth stage deserves the strongest review.
Imagine a sales operator with an overdue invoice, a contract redline, and a meeting that conflicts with travel. A useful brief should identify its evidence, the proposed next step, and the point at which a person must approve. It might draft an invoice reminder, summarize the contract differences, and suggest meeting times. Preparing those actions is different from sending them.
The public demo illustrates a mobile progress update; its example integrations are not proof that your account already has access.
The diagram is our evaluation workflow, not a screenshot of Duet. For every real task, verify the cited source, inspect omissions, and check whether the agent stopped at the right boundary.
Define connector permissions before the first trial
Anthropic says Duet can draw on more than 6,000 connectors and plugins in the Claude Marketplace. Its examples include Gmail, Slack, QuickBooks, Stripe, Gusto, DocuSign, and HubSpot. A connector being listed does not mean that your specific account has granted access, that every integration is available in your region, or that every action is supported. Inspect each app’s authorization screen and the scopes it requests.
Start with three tiers. In a read-only trial, use a test calendar, work notes, or a dedicated low-sensitivity mailbox to judge item discovery. In a draft-only tier, allow customer-mail or agreement summaries but require a person to review every outgoing document. Reserve high-impact systems—payments, payroll, signing, deletion, bulk CRM changes—for a later phase with an owner, a narrow scope, logs, and a rollback path.
Anthropic says Duet waits for approval before sending messages, making payments, signing documents, or deleting anything. Securities trading and transfers between accounts are outside its permitted actions even with approval. Anthropic also says the model does not directly see passwords or sign-in tokens and that additional safeguards limit disclosure of sensitive information. These are vendor claims about design, not a substitute for your organization’s security assessment. Test the visible approval screen, available logs, revocation, and data handling in your own environment.
A small access register is enough to start: app/account | read scope | allowed draft actions | human-approved actions | owner who can revoke. Record a reason for every expansion of access. If Duet pulls information from a source that should not be in that task, disconnect the app and inspect the scope before changing the prompt.
Separate reading, drafting, and execution
The common mistake is not merely connecting too many apps. It is assuming that permission to read a record also means permission to act on it. Reading a test mailbox to identify customer questions is distinct from sending customers a response. Summarizing invoice status is distinct from paying an invoice. Use a task-level matrix before connecting an app:
| Workflow | Read | Prepare without sending | Review before execution | Withhold in first trial |
|---|---|---|---|---|
| Calendar conflicts | Times and participants in a test calendar | Conflict summary and alternative slots | Send a changed invitation | Edit an executive shared calendar |
| Customer email | Messages under a specified label | Reply draft and attachment list | Recipient, body, attachments, then send | Bulk mail and auto-forwarding |
| Invoice reminder | Dates and amounts in a test ledger | Reminder draft and exception list | Send externally or update the ledger | Payment and bank access |
This is an editorial trial aid, not a built-in Duet settings page. The actual permission granularity depends on the connector authorization screen.
Connect only one app at a time. Record who granted access, the intended task, and when it should be revoked. That makes a failure attributable to a specific account and scope. Agree on a stop condition before starting: if a source cannot be explained, internal material appears in an outgoing draft, or a proposed action cannot be fully previewed, pause that connector rather than asking the agent to keep trying.
A useful task instruction has a stopping point
“Handle my work today” is too vague to test. It says nothing about the source, deliverable, or action boundary. A more useful instruction is:
Read only the test calendar for the next seven days. Find overlapping meetings and location conflicts, ordered by start time. For each, provide the original event title, the reason it conflicts, and two possible new slots. Produce a table first. Do not edit the calendar or send invitations. If another account or participant needs to be contacted, explain why and wait for my approval.
After the run, verify that the cited event matches the original record, the alternative is feasible, and the calendar remained unchanged. Run the same task once with the normal manual process; comparing against no baseline makes every fluent draft look like a productivity gain.
Background work still consumes plan limits
In the research preview, Duet is included in eligible US Pro and Max plans but has no separate free allowance. Its checks count against the same usage limit as the rest of Claude; larger tasks such as a financial model or deck require more model turns. Anthropic has not published a stable “cost per morning brief.” It says the several-hour check interval is partly a tradeoff against using plan limits too quickly. Avoid claiming a fixed daily capacity or zero incremental cost.
Measure cost per useful outcome instead of a vague sense that the agent feels active. In one week, count messages that saved a missed step, time spent reviewing and correcting, false alarms, omissions, and whether regular Claude work felt constrained. One simple score is useful alert rate = helpful proactive alerts / all proactive alerts. Another is net time saved = search and preparation time avoided - review and correction time. Record any attempted action beyond the intended scope separately; a high completion count cannot cancel that risk.
Suppose a hypothetical week has 20 proactive alerts, eight of which are useful. Reviewing and correcting them takes 45 minutes, while avoided searching and preparation save an estimated 70 minutes. The useful-alert rate is 40%, and net time saved is only 25 minutes. These numbers demonstrate the calculation; they are not Duet test results. Decide in advance what would justify continued use, such as a minimum net time saving and zero unauthorized external actions. High-impact permission failures should be disqualifying even when the time score looks positive.
Keep one row per day with six columns: date | source task | alert or draft | human verification | review minutes | next action. Record missed items as well; counting only alerts the agent sent systematically overstates its accuracy. Keep existing notifications and your normal to-do list as the control during the pilot.
This is our sample record sheet, not a Duet usage dashboard. The numbers are illustrative placeholders for your own pilot results.
The public morning-brief demonstration shows the alert format. In a trial, evaluate accuracy and review effort rather than notification volume alone.
A seven-day pilot for reversible work
Day 1 — Choose a boundary. Connect one low-sensitivity test account. Pick a calendar-conflict alert or a non-sensitive email draft, not both. Write down forbidden folders, contacts, and actions requiring approval.
Days 2–3 — Establish a baseline. Do the same task using your current workflow. Record time from discovery to completion, errors, and what evidence you used. Do not credit the agent for work nobody would otherwise have done.
Days 4–5 — Observe proactivity. Let Duet produce its brief while keeping existing notifications as a control. For each item, check the source, priority, and timing. At any send, payment, or signing step, confirm that the exact proposed action is visible and awaits approval.
Day 6 — Rehearse a failure. In a test account with fictional data, give an ambiguous instruction such as “handle the customer emails.” Does it ask for clarification or create drafts? Does it attempt to send? This reveals more about the permission boundary than a polished demo task.
Day 7 — Expand or roll back. Add a second connector only if the useful-alert rate, net time, and permission log are acceptable. If a message is sent unexpectedly, a payment action is mishandled, or a source cannot be explained, pause expansion and investigate. Write a narrow conclusion: “Useful for this task under these permissions,” not “Duet works for the whole business.”
Three trial tasks and what each one proves
Calendar conflict. Put two partly overlapping meetings and one nearby but nonconflicting meeting in a test calendar. A good result identifies the first two without flagging the third. Check time zone, location, and attendee interpretation. The deliverable is a suggested change; do not authorize automatic invitations in the first round.
Overdue invoice reminder. Prepare one genuinely unpaid test invoice and one already settled invoice with an old reminder email. Ask for evidence and a draft message. If it treats the old email as proof that the settled invoice is still unpaid, record that failure. Do not give the trial real payment privileges.
Client meeting brief. Ask it to use a dedicated folder of recent communication and a meeting invitation to produce three key points, two questions, and links to the source files. Open a sample of the links. Did it mix information from different clients? Automatic preparation may be useful; sending, editing CRM records, and adding contacts remain review steps.
These tasks cover anomaly detection, communication preparation, and context assembly while remaining reversible in a test environment. Payroll, contract signing, and mass customer notifications are poor first-round stress tests because failures are harder to undo and explain.
How to roll back after a failed trial
If you see an unauthorized action, accidental send, or repeated misjudgment, stop the task and disconnect the relevant connector. Preserve the necessary task record for review, then check external messages, calendar changes, ledger entries, and newly created files. Restore what can be restored, and let the responsible owner handle any communication that has already gone outside the organization. Finally, narrow the read scope, add an approval gate, or mark that task unsuitable for an agent. Deleting the chat is not the same as revoking access.
If you do not have access yet, use the same plan to agree on scopes and evaluation criteria before connecting production data.
Who should wait, and who should watch from a distance?
Duet is most promising for a professional or small-business owner who switches between several apps, checks recurring items daily, and can personally review consequential actions. The potential gain is fewer missed details and less preparation work. That is a testable hypothesis, not a general productivity result.
For real-time incident response, multi-region rollout, centrally managed enterprise permissions, or contractual service-level guarantees, a research preview is not a replacement for established systems. Anthropic says personal errands such as booking tickets are not yet a strength and the preview cannot currently show its browser. Keep critical alerts in dedicated monitoring systems, and test Duet on reversible, reviewable work.
Sources and verification standard
Product capabilities, access, approval boundaries, and usage come from the official Claude Duet announcement, product page and FAQ, and Claude release notes. The pilot plan, scoring, and access tiers are editorial recommendations. We did not independently measure Duet’s performance. Verify current eligibility and controls in your account before deployment.
FAQ
Is Claude Duet available to everyone now?
No. As of October 22, 2026, it is a waitlisted research preview initially for US Pro and Max subscribers. Team and Enterprise plans are not included yet.
Can Duet send email or pay without my approval?
Anthropic says sending a message, making a payment, signing a document, and deleting anything require approval. Verify the exact approval screen with a low-privilege test account rather than treating the announcement as your own audit.
Do background checks have a separate price?
Anthropic has not published a separate per-check price. Preview access is included in eligible US Pro and Max plans, and checks share the plan’s usage limits with ordinary Claude use.
Can it replace real-time monitoring or on-call alerts?
It should not. Duet checks connected accounts every few hours and may misjudge urgency or miss an event between checks. Keep production alerts in a dedicated monitoring and escalation system.
What is the safest first task?
Use a reversible, low-sensitivity task with an easy baseline, such as finding a conflict in a test calendar. Check sources, priority, and approval points before adding email or more connectors.
