Claude Duet is a proactive agent for marketers. Test it before you trust it

Claude Duet is Anthropic's first agent built to work when you are not prompting it, and the question for a marketer is not whether it is impressive but what it can be trusted to do unsupervised. On Anthropic's own pages, Duet has inherited memory, a running thread of what it has done, and an approval gate on sensitive actions. It does not have a documented way to hand it a KPI.

Anthropic opened Duet on October 21 as a research preview by waitlist, starting with US Pro and Max subscribers. We have not had access, so everything below is a reading of public documentation and early coverage, not a hands-on review. Where something is not documented, we say so.

What Duet actually does

Duet sends a morning brief, checks your connected apps every few hours, and messages you only when something needs attention. It runs on its own virtual machine with a browser, so your laptop and phone do not need to stay on. It works with more than 6,000 connectors and plugins in the Claude Marketplace, including HubSpot, Shopify, Stripe, Slack, Gmail and LinkedIn, and it can buy items through Stripe's Link wallet using a virtual card.

The control model is the part to copy. Duet reads and drafts freely, then shows you planned sends, payments, document signings and deletions and waits for your approval. Trading securities and moving money between accounts are blocked even with approval.

Does it have memory, task logs and KPIs?

Here is what the product page and developer Flavio Copes's write-up say as of October 22. Copes states he has not tested Duet, so his points are reading, not observation.

• Memory: yes, inherited. Anthropic says your memory, skills and apps from the Claude app are there from the first message, and that Duet “remembers what you said before.” Copes reports that topic channels share that memory, and that disconnecting an app stops future access but Duet keeps what it already learned. Anthropic does not say how to view or edit what Duet remembers.
• Task visibility: partly. Anthropic says “You can see what Claude Duet has done and is doing for you, and steer it,” and that your history lives in one thread. Copes describes a Tasks area that tracks work and requests for input, where only you can mark a task completed. Neither describes an audit-grade log with timestamps, inputs and outputs for each action.
• KPIs: not documented. Neither source mentions goals, targets or progress tracking. The recurring behavior described is the morning brief and the every-few-hours checks. Whether you can create your own schedules is unclear, and Copes advises confirming what Duet actually set up rather than assuming.

So you can delegate a task with a finish line. On the evidence available, you cannot assign Duet an outcome such as “keep cost per lead under a set number” and have it own that metric. Until Anthropic documents otherwise, write the KPI into the task yourself (metric, threshold, reporting day) and keep your own record of what it reports.

Caveats before you rely on it

• Anthropic's own known limitations: it may misread threads or misjudge urgency, it is weaker on personal errands such as booking tickets, it cannot yet show its browser, and it lives in its own app separate from the main Claude app.
• Timing: checks happen every few hours, so an urgent message can arrive between checks. It is background assistance, not a pager.
• Access and limits: US Pro and Max waitlist only, with Team and Enterprise not included. Background checks and larger tasks count against your plan's usage limits, so an always-on agent can eat your allowance.
• Gaps Copes lists: Microsoft 365 is not available yet, some connectors such as Slack have limits, and Link payments do not support recurring purchases and need a US payment method.
• Shared memory: because Duet inherits your Claude memory, personal and client context sit together. Our recommendation is a work-only account for any test that touches client data.

A one-week test plan

1. Connect the minimum. Start with one read-heavy app, your email or your CRM. Add Stripe, Shopify or anything that can spend or send only after step 4.
2. Days 1 and 2: observe the brief only. Each morning, write down your own top three priorities before you open Duet's brief. Score how many it matched and what it missed.
3. Days 3 and 4: delegate one reversible task with a finish line. For example: every Friday, summarize last week's campaign results from the CRM into five bullets, name the source of each number, and flag anything you could not verify. Check every number against the source.
4. Day 5: test the gate. Ask for something that needs approval, such as an email to yourself. Confirm it stops, shows the exact action, and sends nothing until you approve.
5. Keep a scorecard all week. For each alert, mark it useful, late or wrong. Count missed items, hours saved, and how much of your usage limit the background checks consumed.
6. Decide against a threshold you set before you started. For example: expand only if most alerts were useful and no action bypassed approval.

What to do about it

• Join the waitlist now and plan a work-only account so client context stays separate from personal memory.
• Write the KPI into every delegated task: metric, threshold, reporting day, and what to do when it cannot verify a number.
• Write your approval rules before you connect anything that can send or spend, and copy Duet's read-and-draft-freely, approve-to-act model into every other agent you run.
• Keep your own task log in a doc or sheet until Anthropic documents a real activity record. If you cannot audit it, you cannot delegate to it.
• Treat usage as a budget line. Track what the background checks cost in plan allowance before you add more connectors.