Claude Duet Review: Useful at Connecting Dots, Still Bad at Restraint

Introduction

AI agents are moving beyond coding tools and into everyday office work. Anthropic’s Claude Duet is aimed at professionals and small-business owners who want help managing information rather than writing software. It is designed to connect workplace services, identify what may need attention, and offer to move tasks forward.

Duet remains a research preview, with access limited to certain paid subscribers in the United States through a waitlist. Users can connect services including email, calendars, Google Docs, and Slack. The agent then presents suggestions as swipeable cards that can be accepted or dismissed.

The value comes from connecting services

Duet’s most interesting feature is not text generation by itself. Its advantage is the ability to look for relationships across different sources of workplace information. In testing, it connected details from an automatically generated meeting summary with a Slack message and a shared document. It also inferred aspects of the team structure from recurring work patterns.

Each morning, Duet provided a short overview of the day and surfaced practical reminders, such as preparing an item before a work trip or shoot. It could offer to add an agenda item, draft a message, or update a shared document. When planning several weeks of work, it identified that upcoming time off would leave very little time before another major draft was due. It then proposed an adjustment, prepared weekly notes for an editor, and drafted a Slack update.

That kind of cross-application checking may be more valuable than a conventional chatbot for people who struggle with scattered commitments. Modern work is distributed across inboxes, direct messages, meetings, and documents. An agent that can compare those sources and flag inconsistencies has a practical role.

Helpful, but sometimes overly anxious

The same proactivity can become irritating. In one example, Duet warned early in the morning that an article was still awaiting edits, apparently because that could interfere with a promised link. The warning missed the nuance that the stated publication day was not a firm deadline and that the editor’s local time mattered.

This points to a broader limitation. Duet can identify a dependency, but it does not always understand its importance, social context, or timing. It behaves like an assistant that is very attentive to unfinished tasks but not yet skilled at deciding which tasks can safely wait.

Its drafted messages can be impressively tailored to a user’s habits, including familiar capitalization and workplace abbreviations. Yet it may also introduce wording the user would never choose. Messages sent with Duet are labeled as such, so the system is not a completely invisible substitute for human communication.

A backstop, not a creative partner

The clearest boundary concerns creative work. The reviewer did not want Duet to start drafts or review its own contribution. Writing involves more than producing acceptable language: it also includes finding unexpected connections, developing a personal angle, and making choices through the process itself.

That makes a useful distinction between administrative assistance and creative delegation. Duet is strongest when it reminds, cross-checks, and prepares low-stakes communication. It is less convincing when it tries to take ownership of the thinking that gives work its individual character.

Why it matters

Duet also makes the privacy trade-off more visible. The agent becomes more useful as it receives access to more workplace data, but that access can make users feel continuously observed. Offloading calendars, to-do lists, and meeting follow-up may reduce friction while also creating distance from the work itself.

Claude Duet is therefore best understood as an information integrator and organizational safety net, rather than a fully reliable personal assistant. Its cross-service awareness is genuinely useful, but the next step is not simply more access. It needs better judgment about urgency, tone, local context, and when not to interrupt.

Source: The Verge AI