Anthropic has introduced Claude Duet, a proactive agent for professionals and small business owners that can continue working after they close their devices. Announced on October 21, the product enters a research preview for U.S. Claude Pro and Max subscribers through a waitlist.
The company says Duet can use more than 6,000 connectors and plugins in the Claude Marketplace. That is a combined catalogue figure, rather than a count of 6,000 distinct applications or evidence that every connected workflow performs equally well.
For businesses, the practical question is what happens between an agent noticing work and receiving permission to carry it out. A useful assistant needs to surface the right issue, prepare an accurate next step and recognise when a person should decide.
Practical guide: For current eligibility, subscription pricing, setup, and preview restrictions, read Claude Duet explained: access, pricing, features, and limits.
What Claude Duet does in the background
In its launch announcement, Anthropic describes a cloud virtual machine with its own browser. Duet can keep working without the user’s phone or laptop remaining on, bringing across memory, skills and connected apps from their Claude account.
Its proactive monitoring has a cadence: connected accounts are checked every few hours. “Always-on” therefore does not mean continuous observation of every incoming event.
Anthropic’s examples include locating unpaid invoices in QuickBooks or Stripe, preparing follow-ups from HubSpot records and helping with agreements through Docusign. These are announced workflows, with their usefulness depending on the information and access available in each connected account.
Consider a hypothetical overdue invoice. Finding it is one task. Checking whether a customer has already disputed it, drafting a suitable reminder and choosing the correct recipient require additional context. Connecting the accounting system and inbox would make that context available, but would not by itself establish that the proposed reminder is appropriate.
Access and background usage have limits
Duet runs on Opus 5.5 and is offered through separate iOS and Android apps and the web. Reuters reported the launch as Anthropic’s entry into a growing market for proactive assistants competing with offerings from OpenAI and Meta.
The preview is included in eligible subscriptions and shares their usage allowance with the rest of Claude. Background checks consume that allowance, while larger assignments require additional model activity. Team and Enterprise plans are excluded from the initial preview.
That creates an operational trade-off. A business owner asking the agent to watch several accounts and produce a substantial report is allocating capacity across both activities. The relevant cost is partly the subscription and partly how much useful work remains possible within its limits.
Anthropic also acknowledges that Duet can misunderstand a conversation, misjudge urgency or miss developments between checks. Those limitations matter particularly for time-sensitive work: an assistant’s silence should not automatically be interpreted as evidence that nothing needs attention.
Approval remains a separate step
Anthropic says Duet waits for permission before sending messages, making payments, signing documents or deleting material. Securities trading and transfers between accounts remain prohibited even with approval.
The distinction between preparing and authorising an action also appears in Visa, BILL and Billtrust’s supplier-payment pilot, where agents execute payments buyers have already approved.
For a reviewer, an approval request is useful only if the proposed action is understandable. A payment prompt should make the recipient and amount clear. A proposed message should expose the actual wording and destination. Otherwise, a human can remain formally involved while having too little information to catch a mistake.
This is an evaluation principle, rather than evidence that Duet’s approval prompts succeed or fail in practice.
The test is whether proactive help reduces work
A meaningful trial would measure the effort required to review and correct the agent alongside the work it completes. A morning briefing that looks polished can still create extra work if important items are omitted or routine messages repeatedly become urgent alerts.
A business could evaluate a bounded workflow using known examples, checking missed deadlines, unnecessary interruptions, draft corrections and usage consumed. These would help establish whether the assistant earns responsibility for a wider set of tasks.
Duet’s launch puts those questions into an everyday business setting. The catalogue establishes the potential reach; repeated, accurate work within a clear approval boundary will determine how much responsibility users can comfortably delegate.
Prepared with AI-assisted research and drafting. AgentsCalling has not independently tested Claude Duet.
About Puneet Sharma
Puneet Sharma · Freelance Writer
Puneet Sharma is a freelance writer covering AI agents, automation and the software changing how we work. His coverage at AgentsCalling starts with the questions a reader might ask over a conversation: What does this tool actually do? Where does it fall short? And is it useful beyond the demo? His articles explore new launches, practical workflows and the trade-offs behind giving software more responsibility, with explanations that make room for readers who are still getting familiar with AI.
