Claude Duet tests the proactive assistant, not a new chatbot
Anthropic’s Claude Duet preview points at a useful shift: AI that watches, filters, drafts, and asks for approval before acting. The promise is less chat and more operational triage, but the hard part is trust, scope, and error recovery.
TL;DR: Claude Duet is interesting because it moves Claude from “answer my prompt” toward “notice what needs my attention,” but the product only works if approval, memory, and connected-app permissions are handled carefully.
What is Claude Duet actually trying to change?
Anthropic’s primary source here is its video, “Meet Claude Duet,” where Josh from Anthropic describes an early research preview of a more proactive Claude experience.
The notable shift is not that Claude can draft an email or make a plan. Plenty of assistants can do that now. The shift is that Claude Duet “messages me first,” as Josh puts it, before he knows what to ask. In the demo, it filters incoming video requests into a short morning list, flags an email he meant to answer, drafts a reply about a back-ordered camera, and turns one long voice note into a shoot plan and gear list.
That sounds small. It is not.
Most AI work still starts with the user doing the hardest part: noticing the problem, deciding it matters, gathering context, and writing a prompt. Duet is pointed at the step before the prompt. It is trying to become an attention router.
Anthropic also says the same Claude memory and connected apps carry over, so the user does not have to “explain myself all over again.” That matters because a proactive assistant without memory is just a nagging notification system. A proactive assistant with too much authority is a liability.
The design choice Anthropic emphasizes is approval. “Nothing ever gets sent out until I hit the approve button,” Josh says. That is the right default. Not glamorous, but right.
Does this prove the chief-of-staff agent is here?
No. It proves Anthropic is testing a shape of product that many teams want.
The “personal chief of staff” framing is catchy, but a real chief of staff does more than summarize tasks. They understand politics, tradeoffs, timing, priorities, and consequences. Claude Duet, as shown, is closer to an executive assistant for bounded workflows: triage this inbox, draft this reply, plan this shoot, remind me about the task I dropped.
That is still useful.
The key question is whether it can stay useful without becoming noisy. Proactive AI has an obvious failure mode: it interrupts you with confident nonsense, half-prioritized tasks, or stale context. If it misses something urgent, trust drops. If it over-flags everything, users turn it off.
The demo gives good examples because they are concrete. A back-order email. A studio booking. A gear list. A payroll reminder. Unpaid invoices. These are workflows with known artifacts and obvious approval points. That is where this kind of product should start.
Not strategy. Not “run my company.” Start with recurring operational friction where a draft is valuable and a human approve button is cheap.
Anthropic says Claude Duet is an early research preview, and users on Pro or Max in the US can join the waitlist. That is all I would treat as confirmed from the first-party material. No pricing changes, timelines, or broader availability were provided in the source.
What should builders test first?
If you are building with this pattern, do not start by asking, “What can the agent do?” Start with, “What should the agent notice?”
That changes the product spec.
A useful Duet-style assistant needs a watchlist, not just a prompt box. It needs to know which inboxes, calendars, docs, CRMs, payment systems, ticket queues, or project boards matter. It needs a threshold for interrupting. It needs a memory of past preferences, but also a way to forget bad assumptions. It needs dry runs, audit trails, and obvious human approval before any external action.
The first version should probably do three things: identify items that need human attention, prepare the next action, and explain why it surfaced them. If the explanation is weak, the user will not trust the recommendation. If the action is hidden, the user will not trust the system.
My practitioner’s take: pick one recurring workflow where you already lose time to triage, then build a “suggest-only” agent around it. Let it read the relevant context, return a short priority queue, draft the next step, and require approval every time. The catch most teams miss is that the hardest product decision is not model choice. It is deciding when the assistant is allowed to bother you.