Claude Duet: what to delegate, what to check and how to start

What is Claude Duet, and what changes when an agent is proactive?

Duet is designed to follow work through, rather than wait for every question. Give it a goal and approved context, and it can prepare a result or bring back a decision. The practical difference is not simply better writing. It is whether a useful piece of work is ready when you need it.

Consider a sales meeting tomorrow. A useful assistant pulls together the latest proposal, previous commitments and unanswered questions. An unhelpful one summarizes twenty emails without identifying what needs a decision. Both produce text. Only one reduces the work you still need to do.

Duet is not a new foundation model or another name for Claude Code

Anthropic makes both products, but the jobs differ. Duet is an ongoing assistant across connected work apps. Claude Code is aimed at working with code and projects. Having a cloud computer does not give Duet automatic access to every file or development session on your laptop.

For repository changes, review and version control, our Claude Code guide describes a different starting point. ChatGPT Dots belongs to the broader always-on agent category too. Compare the actual workflow and permissions, rather than assuming that similar positioning means identical capabilities.

Who can use Claude Duet right now?

The announced research preview is for invited US users with Pro or Max. As of October 30, 2026, it is not a worldwide release, and a subscription alone does not secure an invitation.

Access is offered through standalone iOS and Android apps and the web. Follow the invitation associated with your own Claude account. Do not treat an unofficial site using the Duet name as an Anthropic service or pay a third party for guaranteed access.

If your market is not eligible, there is still useful preparation to do: pick one recurring task, identify its minimum data requirements and create an example of a good result. That work remains useful regardless of which agent eventually fits your business.

How much does Claude Duet cost?

During the preview, eligible users receive Duet within Pro or Max without a separate Duet fee. US list pricing is $20 for a monthly Pro subscription or $200 billed annually. Max has $100 and $200 monthly tiers. Taxes and local billing can differ.

The hidden planning issue is capacity, not a hidden Duet subscription. Background work counts against the same plan limits as your other Claude activity. Morning briefs and recurring checks can use capacity even when you are not actively chatting. Always available to work does not mean unlimited work.

Measure cost per accepted result

A twenty-page report is not a saving if correcting it consumes an hour. Include setup, review and rework when deciding whether a task is worth delegating. Start with a low check frequency and a small set of inputs, then expand only when the results and remaining plan capacity justify it.

How does a Duet task move from request to result?

A practical assignment combines a defined outcome, approved app context and a review point. Duet can take typed or spoken requests and continue on its cloud computer. What it can use still depends on which information is actually accessible and up to date.

Connected work apps include Gmail, Google Calendar, Google Drive, Slack, HubSpot and QuickBooks. Tasks shows ongoing work and decisions; created documents and presentations appear in Make. A connection provides context, not blanket permission to send every message or change every record.

1. Define the outcome. Specify an artifact, deadline and boundaries. “Prepare three questions for this client meeting” is easier to evaluate than “manage my marketing”.
2. Supply the smallest useful context. Use approved files or necessary app connections. Not every task needs access to an entire mailbox.
3. Let it prepare and follow up. The agent can return with a draft, a file or a decision. Account checks run at intervals rather than as continuous monitoring.
4. Inspect the deliverable. Open the actual file and verify the evidence and task status, rather than accepting a chat message saying it is finished.
5. Approve a specific next action. Match the recipient, content and any amount against your original request before an external action proceeds.

Keep urgent work on a separate escalation path. A critical message can arrive between checks. Also define when the agent should stay quiet: repeated reminders about an unchanged task can create more work than they remove.

Which business tasks are worth trying first?

Start with preparation that can be checked before it affects a customer. These are proposed assignments, not measured Jakency outcomes or promises that every connector supports every action.

Meeting preparation and internal analysis are sensible first pilots because a mistake can be caught within the team. Payments, contracts and outbound communications deserve tighter controls. If the task always follows the same deterministic rules, a purpose-built AI automation may be a better fit than having an agent reason through it again each time.

How should you set up your first Duet assignment?

Once invited, start with one bounded job. Review existing Claude app connections before adding more: previously connected context may already be available.

1. Use the invitation for your account. The waitlist email must match your Claude account. Follow the device setup provided with that invitation.
2. Choose one workflow. Keep meeting briefs separate from payroll, personal errands and payments during the first test.
3. Limit inputs. Start with a small approved file set or shared folder. Check what a connector can read and change.
4. Define acceptance and stop conditions. Specify what the result must contain and when missing information should trigger a question.
5. Run once before scheduling. Compare the output with a manually checked example before allowing recurring activity.

Goal: prepare tomorrow's client meeting.
Use only the supplied calendar entry, proposal and meeting notes.
Return no more than 1 page: objective, confirmed commitments,
3 open questions and a proposed next step.
Identify the source file and date behind each factual statement.
Flag missing or conflicting information.
Do not send messages, change appointments or connect more apps.
Ask before making a decision the evidence does not support.
Deliver a draft for my review before use.

“Sounds professional” is a weak acceptance test. “Uses the current proposal and makes no unapproved pricing commitment” is one you can actually check. Keep the original instructions alongside the deliverable so a colleague can apply the same standard.

How can Claude Duet help with SEO?

Use it first to analyze supplied evidence, not to publish automatically. A useful SEO review ends with a justified action on an existing page, a request for further investigation or a decision to leave things alone. It should not turn every keyword into a new URL.

Example: turn Search Console exports into five decisions

Provide two comparable periods with consistent country, device and search-type filters. Include query-page pairs where available. Separate keyword and page tabs do not establish which query belongs to which URL. Add your current page inventory and each page’s intended job.

Review these two GSC periods and our existing URL inventory.
First check dates, filters, columns and missing data.
Calculate CTR as clicks / impressions, not an average of row CTRs.
Return no more than 5 recommendations, each with:
query intent, existing URL, measured signal, evidence row,
possible explanation and the next verification step.
Separate observations from hypotheses.
Do not label multiple ranking URLs as cannibalization by default.
Do not invent volumes, positions or relationships between separate tabs.
Require an intent and SERP check before suggesting a new page.
Make no site changes. Publish nothing. Return an auditable proposal.

A simple calculation illustrates the boundary: 12 clicks from 600 impressions is a 2% CTR. An agent can calculate that. It cannot establish from those two numbers that a different title will generate more customers. Search intent, positions and conversion evidence are still missing.

The local connection described in our Search Console MCP guide for Claude is not automatically supported in Duet. An approved CSV export is a clear starting point. The decision still needs keyword research and a review of current search results. For citations in AI answers, use a separate AI visibility measurement process. Using an agent does not itself make your business a cited source.

What do the early demonstrations tell us?

The supplied demonstrations make task organization, briefing documents and voice-directed presentation edits tangible. They do not establish a general productivity gain or reliability rate.

Riley Brown’s demonstration shows a central conversation, topic channels and separate task threads. He likes editing presentations from his phone, but describes the additional conversation layers as potentially confusing. Another reviewer reports that completed work did not always carry clearly into another channel. These are individual experiences of an early version, not a controlled benchmark.

Turn that into an acceptance test: can you tell which task produced a file, which version is current and where completion is recorded? “Done” in the conversation is not enough if another task is about to send the same follow-up again.

Claims that an agent “builds a business”, “replaces an employee” or works “ten times faster” need more evidence than an impressive demo. A prototype is not a paying customer, and a polished campaign draft is not proof of profitable delivery.

What permissions and limitations should you account for?

Keep access proportional to the job. Anthropic provides approval and safety mechanisms, but you still need to verify recipients, amounts and information. An explicit instruction may itself provide consent: “send this” and “draft this” are different authorizations.

• Untrusted instructions. Email and web content can contain prompt injection. Material the agent reads should not become permission to take unrelated actions.
• Persistent context. Disconnecting an app stops new access, but does not automatically remove what was already retained in memory. Review chat and memory controls separately.
• Stale information. Scheduled checks are not continuous monitoring. Recheck the current state before sending a reminder or approving a consequential action.
• Connector differences. Microsoft 365 is not yet available in Duet, and other connectors can have limitations. A listing elsewhere in Claude does not prove feature parity.
• Local files. Do not assume that the cloud agent can see all projects and files on your computer. Supply the necessary material deliberately.

The preview also does not let you watch or take over Duet’s cloud browser. For important actions, verify the outcome in the relevant app itself. Only introduce confidential information when the account settings and your organization’s data rules permit it.

Write down three categories: preparation it may do alone, actions requiring explicit approval and work it must never do. That is more usable than a general instruction to “be responsible”. Our work on AI for businesses starts with those operational boundaries.

How do you evaluate Duet in seven days?

Compare the same type of task with and without the agent, including review and repair. A short pilot supports a decision about that workflow, not a claim that the whole business is more productive.

1. Day 1: baseline. Perform a representative task manually. Record time, inputs and acceptance criteria.
2. Day 2: draft run. Give the agent a comparable task using a limited, approved dataset.
3. Days 3 and 4: exceptions. Test a missing file, conflicting dates and an already completed task. Does it ask, stop or invent an answer?
4. Day 5: handoff. Have a colleague use the output without reading the entire conversation.
5. Days 6 and 7: decision. Include review time. Keep, narrow or stop the workflow based on accepted quality and net time saved.

Illustrative calculation: a manual brief takes 35 minutes. With Duet, instructions take 4 minutes, review takes 9 and corrections take 7. Net saving: 35 - 4 - 9 - 7 = 15 minutes. If corrections take 25 minutes, the workflow instead costs 3 extra minutes. Generation speed alone hides that difference.

Our recommendation is to treat Duet as a candidate for work that falls between apps and people. Start with one deliverable you can judge, grant only the necessary access and expand after the results hold up. Outside the US preview, prepare the workflow now without assuming you can deploy the product today.

Frequently asked questions about Claude Duet