I call my ChatGPT dot Sam. One of the features I like most is being able to call and speak to it, then ask for help with projects running in Codex on my MacBook at home while I’m away. The useful connection is to work I can pick up again, even when I’m nowhere near that keyboard.
That experience gives Claude Duet , Anthropic’s newly announced proactive agent , a concrete point of comparison. Can it carry a job forward between conversations? Can I steer it without reopening every app? And does it know which decisions need to come back to me?
Anthropic opened the research preview on October 21. Duet is built around Claude Opus 5.5 and aimed at professionals and small-business owners. Its announcement is a promise to examine through that working relationship. My experience here is with Sam, while Duet’s capabilities below come from Anthropic’s documentation and the launch reporting.
The work continues on Duet’s own computer
Duet has its own cloud computer and browser. That is how it can keep working when a user’s phone or laptop is switched off. Mashable’s launch report describes the practical range: tracking messages and tasks, preparing business work and returning when something is finished or needs a decision.
The continuity is appealing. Anthropic says Duet carries over memory, skills and connected apps from an existing Claude account. Its product page describes one ongoing thread, with text or voice as the way to hand over a task. A morning briefing and later check-ins give that thread a reason to resume without another opening prompt.
TechCrunch places Duet alongside agents including OpenAI’s Dots and Meta’s Muse, while emphasizing its work and small-business focus. Scheduling, dashboards and forecasts are among the announced uses. These are launch descriptions, rather than independent measurements of how dependably Duet completes them.
What calling Sam adds to the comparison
In my earlier article about OpenAI Dots, I looked at ongoing assignments: following a changing brief, preparing revisions and returning with something a person can inspect. Sam makes that idea personal. I can talk through what I want and ask for help with Codex projects on the MacBook I left at home.
There are two useful connections in that experience. One is conversational, the ability to call and speak. The other reaches the environment where my projects already run. They solve different parts of being away from my desk.
For Duet, Anthropic documents voice messages and a separate cloud machine. That does not establish the same live-call experience or access to the Codex projects on my own Mac. I would want to understand those details before expecting my existing routine to transfer. A cloud computer can be valuable while still leaving a different handoff between the assistant and a local project.
A useful first assignment would therefore be specific. For example, ask an agent to follow one approved project brief and prepare proposed edits when it changes. Give it a clear end date and ask it to return the changes for review. That makes the result inspectable before entrusting it with a larger continuing job.
Approval belongs beside the action
According to The Verge’s report , Duet seeks approval before major steps such as a payment or deletion. Anthropic’s announcement specifies messages, payments, document signing and deletion as actions that require confirmation. Securities trading and transfers between accounts remain outside its permitted scope, even with approval.
The useful test is what that confirmation shows. For an outgoing message, I want the recipient and final wording together. For a proposed change, I want to see what will change. A vague request to continue gives me less to decide than the actual action waiting in front of me.
Anthropic also describes credential isolation and a classifier intended to prevent sensitive information from being shared improperly. Those protections address a real part of the problem. The agent’s reading access deserves separate attention, because asking before it sends something does not answer every question about the information it has already received.
A privacy caution about Meta Muse
I would approach Meta Muse with particular caution before connecting private messages, personal records or business accounts. Meta’s privacy history makes that a substantive concern, rather than a preference for one assistant’s personality.
In 2019, Facebook agreed to a $5 billion FTC settlement resolving charges that it violated an earlier privacy order by deceiving people about their control over personal information. The FTC described data-sharing practices that undermined users’ choices. That documented history is a reason to examine Meta’s current access, retention and data-use terms closely before handing Muse an intimate view of everyday work.
It does not establish a particular privacy violation by Muse. My caution is about choosing whom to trust with that access. The same questions about what is read, retained and shared belong in any evaluation of Duet or Dots too.
The background work also uses your allowance
The Neuron highlights a less visible part of the bargain: background checks and larger tasks consume the existing Claude allowance. Anthropic says checks run every few hours, with the interval partly reflecting that cost. This is continuing work with a usage budget, not constant monitoring for free.
Access is also limited. Duet is a waitlisted research preview for Pro and Max subscribers in the United States, available through its standalone phone apps and the web. Anthropic says it can misread threads, misjudge urgency and miss something between checks. The launch version cannot yet show its browser, which limits one way of inspecting what it is doing.
For me, the first useful result would be modest: one relevant development, a prepared piece of work and a clear decision to make. Calling Sam already gives me a way to stay involved with projects while I’m away from home. That is the experience I would bring to Duet: let the work continue, but make the next conversation easier to pick up.
