Anthropic introduced Claude Duet on October 21, 2026, as a proactive assistant for professionals and small business owners. It can continue working after a user steps away, then return with completed work or a decision that needs attention. The launch is a research preview, initially offered through a waitlist to US Pro and Max subscribers.
The announcement raises a practical question: can an assistant reduce the effort of keeping work moving without creating another stream of notifications to manage?
What makes Claude Duet different?
According to Anthropic’s product page, Duet uses Opus 5.5 and carries over a user’s Claude memory, skills and connected apps. It can prepare a morning briefing, flag items in an inbox or calendar, and assemble documents, presentations or dashboards.
The significant change is the working relationship. A conventional chat session usually starts when someone remembers to ask a question. A proactive assistant also has to decide when to bring something to the user.
That makes timing part of the product’s quality. An accurate summary delivered after a deadline has little value. An assistant that repeatedly interrupts over minor issues can be equally unhelpful.
Where background work could help
Anthropic’s announcement says Duet runs on a cloud computer with its own browser, so a user’s laptop or phone does not need to remain on. Examples include finding unpaid invoices and drafting reminders, preparing meetings and building business dashboards. Connected accounts are checked every few hours rather than continuously.
Consider an illustrative workflow for a small agency: before a client meeting, an assistant could gather the latest correspondence, identify unanswered questions and prepare a short briefing. The useful result would be a clear account of what needs a decision, with links back to the supporting material.
Success should be measured against the whole workflow. Gathering information quickly helps only if reviewing and correcting the briefing takes less time than preparing it manually.
Our guide to what you can actually delegate to AI agents provides a framework for choosing a manageable first task.
How the approval controls work
In its October 23 safety explanation, Anthropic describes approval requests that show the proposed action, destination and content. Important steps, such as payments or potentially sensitive messages, require user approval. Some activities, including securities trading and moving money between bank accounts, remain prohibited even with permission.
Anthropic also describes a separate automated check that assesses whether an action matches the user’s intent. Another classifier checks outgoing messages for sensitive information.
For a user evaluating the preview, a useful question is whether each approval contains enough context to make a decision. A reminder may look reasonable while naming the wrong customer or referring to an invoice already paid. Review needs to cover the underlying facts as well as the wording.
Why keeping credentials separate matters
Anthropic says Duet’s model cannot directly access passwords, browser cookies or connected-app sign-in tokens. It also says disconnecting an app stops new access, while information already retained in memory remains until that memory is reset.
The credential design follows an approach explained in Anthropic’s Managed Agents engineering article: separate the system coordinating work from its execution environment, and keep service credentials outside that environment. A proxy can authenticate a tool request without handing the underlying secret to the agent.
This distinction matters because permission to use a service and possession of its credentials create different risks. Keeping a token outside the working environment reduces one route through which generated code or malicious instructions could expose it. It does not establish that every permitted action will be appropriate.
What remains uncertain
Browser agents encounter information supplied by other people. Anthropic’s browser-security research explains how attackers can hide instructions in that material to redirect an agent. The company explicitly treats prompt injection as an unresolved problem.
Its Duet safety documentation also acknowledges that sensitive-information checks can miss problems or flag harmless content. These are the vendor’s descriptions of its protections, rather than independent evidence of reliability across every business workflow.
Availability and usage impose further limits. The product FAQ lists US Pro and Max subscribers on a waitlist; Team and Enterprise plans are not included yet. Background checks consume the same plan allowances as other Claude work, and Anthropic warns that Duet may misunderstand a conversation or misjudge urgency.
There is therefore no basis here for assuming immediate access in France or unlimited background automation.
What to measure in a first trial
DemystIA’s assessment is that a useful trial should begin with one recurring task and a clear definition of a satisfactory result. For a meeting brief, that might mean correct participants, current source material, identified open questions and no unsupported conclusions.
Track preparation time saved, corrections required, interruptions generated and usage consumed. Include the time spent reviewing approvals. Our article on the human attention bottleneck explores why that review effort deserves its own measurement.
Claude Duet makes proactive assistance more concrete. Its practical value will depend on whether users receive useful, timely work they can verify with less effort than the task previously required.
Sources
- Anthropic — Introducing Claude Duet, October 21, 2026.
- Anthropic — Claude Duet product page and FAQ, checked October 27, 2026.
- Anthropic — Claude Duet Safety, Privacy, and Security, October 23, 2026.
- Anthropic — Scaling Managed Agents: Decoupling the brain from the hands, April 8, 2026.
- Anthropic — Mitigating the risk of prompt injections in browser use, November 24, 2025.