Claude Duet for Product Managers: Anthropic's Ambient Background Agent

Anthropic shipped two distinct agentic products in 2026, and the distinction matters. Claude Cowork (launched July 2026) is a task agent: you open a session, give it a goal, and it works through the steps to complete it. Research, drafting, code review, analysis. You trigger it; it acts; it stops.

Claude Duet is different in its fundamental posture. It does not wait for you to invoke it. Duet runs continuously in the background, monitoring your connected apps and surfacing information on its own schedule. The closest analogy is an executive assistant who reads your inbox before you wake up, prepares a morning briefing, flags the three things that need a decision today, and waits for your say-so before taking any action.

The research preview label means Anthropic is collecting structured feedback before broader rollout. Expect feature gaps, known issues, and rapid iteration over the next 60 to 90 days. US-only availability at launch reflects the legal complexity of an agent that can take real-world actions on your behalf: data handling, liability for misread instructions, and authorization scope all become thornier when the AI is proactive rather than reactive.

Duet's architecture has three layers. Understanding them matters because they map directly to the product design decisions you face when building something similar.

The first layer is the connection layer. Duet hooks into your calendar, email, and connected apps through OAuth and Anthropic's connector infrastructure. At launch, connectors include Gmail and Google Calendar. Each connector has a defined read scope: Duet can read events and messages but cannot write without explicit user approval. This scope boundary is not an implementation detail. It is the product's trust model.

The second layer is continuous inference. Duet runs inference against new data in connected apps on a schedule, roughly every 15 to 30 minutes based on observed preview behavior. It is not streaming your inbox in real time. It polls, processes, and updates its internal state. The Opus 5.5 model choice matters here: ambient monitoring is not a simple classification task. It requires the model to maintain context across sessions, understand your priorities from prior interaction, and reason about salience rather than just summarizing.

The third layer is action approval. When Duet identifies something that requires action, it generates a proposed action and presents it to you for confirmation. The categories that require approval are what Anthropic calls the four-S actions: send (email, messages), spend (purchases, subscriptions), sign (commitments, agreements), and delete (irreversible removals). Everything outside these four categories can happen without prompting you. Adding a calendar note does not need approval; sending a reply on your behalf does.

Based on the research preview, Duet's daily surface looks like this. Each morning, it generates a briefing covering your calendar for the day, outstanding email items that need action, and any time-sensitive flags from connected apps. The briefing is delivered as a push notification with a summary card; tapping it opens the full briefing in the claude.ai or mobile app.

Throughout the day, Duet surfaces alerts when something changes. If a meeting is moved, a message from a flagged contact arrives, or a calendar conflict appears, Duet sends a notification and proposes a response if one is warranted. For a rescheduled meeting, it might offer to update your calendar block and send a reply in one confirmation tap.

The approval flow appears as a card with a proposed action and a clear approve or dismiss button. Approved actions execute immediately. Dismissed actions are logged but not retried automatically. Early users in the research preview report that the approval cards are well-designed for quick decisions but that the volume of proposals is still being tuned.

Duet is the most fully realized ambient agent from a major lab released to the public as of October 2026. That makes it an unusually good reference design for AI PMs building agentic products. Three lessons stand out.

Trust is earned through constraint, not capability. Duet is deliberately limited at launch. It cannot send email without approval, cannot make purchases, cannot delete anything. These are not technical limitations; they are product choices. Anthropic is building trust with users before expanding scope. The pattern applies to any agentic product: start with read-only or low-stakes actions, demonstrate reliability, then expand.

Ambient agents need explicit briefing rhythms. Duet anchors its value delivery to the morning briefing rather than interrupting the user throughout the day. This gives users a predictable moment to engage with the agent and sets expectations about when and how it will communicate. Products that try to be constantly present feel intrusive; products with clear rhythms feel like a useful habit.

Approval UX is a product, not a feature. The approval card in Duet needs to convey enough context for the user to make a confident decision in two to three seconds, without reading a paragraph of explanation. Getting this right requires testing dozens of card formats. If your agent needs approval flows, allocate design effort proportional to how often users will see the card. A card users see ten times a day will define whether they trust the agent or turn it off.

The research preview label signals active iteration, not polish. Several limitations affect how useful Duet is today.

Email misclassification is the most-cited issue in early reports. Duet sometimes flags routine emails as urgent and misses genuinely time-sensitive ones. This is a priority calibration problem, not an accuracy problem: the model can read the emails correctly but cannot yet weight them against your specific context well enough to sort them reliably. Anthropic is using preview feedback to improve this.

Usage counts against your plan limits. Because Duet runs Opus 5.5 continuously, it consumes tokens on your Pro or Max plan throughout the day, not just during active sessions. Users on Pro plans report hitting limits faster than expected. Anthropic has acknowledged this and is working on a usage model that separates ambient background consumption from active session use.

The connector set is narrow at launch. Gmail and Google Calendar cover many users, but the absence of Outlook, Slack, and Notion integrations limits Duet's utility for teams whose workflows live outside Google. The roadmap includes these integrations but no dates have been given.

US-only availability limits the preview population and slows iteration for non-US teams. Anthropic has not given a timeline for international availability, citing ongoing review of data handling requirements by jurisdiction.