Anthropic's Claude Duet Pauses for Approval Before Costly Actions
Anthropic has described Claude Duet, a proactive agent powered by Opus 5.5, as a research preview available on a waitlist to Claude Pro and Max subscribers in the US. The company's Claude product page is the primary description of the product.
According to the page, Duet takes on tasks of all sizes and keeps working while the user is away. It waits for the user's approval when a task reaches something that matters. Sending messages and paying are named as examples of sensitive operations that require sign-off. In the design's framing, the model is autonomy by default, with the human pulled back in at the point where an error becomes costly.
The execution details are the more unusual part of the description. Per the product page, Duet runs its approvals, logins, and safety checks in its own cloud computer, a dedicated environment separate from the agent's immediate context, meaning the information the model holds while it carries out a task. For engineers who have built or reviewed agent systems, that places credential handling and policy enforcement in a dedicated environment rather than leaving it to the agent's working context.
Duet is a standalone app for iOS and Android, and it can also be used on the web at claude.ai. Access during the research preview is limited to US subscribers on Pro and Max plans. Team and Enterprise plans were not included in the preview.
Duet shares the user's plan usage limits with the rest of Claude during the preview. Long-running agent sessions therefore draw on the same allowance as ordinary chat and coding work, which matters for anyone planning heavy workloads. The product also works with more than 6,000 connectors and plugins, the add-on integrations that let an agent read from or act in other software.
In my view, the approval checkpoint is where this product will be judged. An agent that acts while the user is away is only as useful as its ability to tell a routine step from a consequential one. Whether that line is drawn sensibly, and whether approval prompts give the user enough context to decide quickly and correctly, will matter more to early users than the headline capability.
The broader context here is a set of concrete questions for teams watching agent products. What counts as a sensitive operation? How is a pending approval presented? How much of an agent's work gets reviewed before it executes, and how much is simply trusted?
The upside, if the approval model holds up under real use, is easy to describe. Tasks that would otherwise sit in a queue until the user returns can move forward in the meantime, and the human keeps the final say on actions that carry consequences. That combination is the point of the design, and the waitlist will show how well it works outside the product page.
