Claude Duet Makes Approval Design the Real Agent Test

Anthropic launched Claude Duet on October 21, 2026 as a research-preview proactive agent app powered by Claude Opus 5.5. The product provides daily morning briefings and executes business tasks in the background across more than 6,000 connectors, according to Anthropic. Sensitive actions such as payments require user approval.

Access is through a waitlist for US Claude Pro and Max subscribers. That matters for buyers interpreting the launch: a research preview with a subscriber gate is not the same proposition as a broadly available enterprise operating system. The immediate opportunity is to examine how proactive work reaches a human decision, rather than assume the integration count establishes deployment readiness.

TechCrunch described Duet as “Anthropic's bet on agents.” The competitive field includes Meta Muse, OpenAI Dots, and Google's Gemini Agent. For a business buyer, the useful distinction is not merely which assistant can start work without a new prompt. It is which product makes the scope of that work understandable before anything consequential happens.

Background work changes the point of control

A conventional request gives the user an obvious moment to define intent. Background work moves that moment away from the eventual action. A morning briefing can be reviewed after it appears; a proposed payment needs a decision before execution. Duet's requirement for approval on sensitive actions acknowledges that difference, but the requirement alone does not describe the entire review experience.

The buyer should ask what information accompanies the approval. A reviewer needs to understand what is about to happen and why the agent considers it appropriate. Those are evaluation questions, not claims about features Anthropic has shipped. The product description establishes an approval boundary; the customer still needs to inspect whether that boundary is intelligible in the context of actual work.

Claude Opus 5.5 is the named model behind Duet. Model capability and authorization remain separate issues. A system can prepare a sensible action without having permission to carry it out. Conversely, permission to access a connected service does not establish that every proposed action in that service is justified. Proactive deployment makes those distinctions part of product design rather than an occasional exception.

Connector breadth is not workflow coverage

Anthropic's 6,000+ connector figure describes a broad integration surface. It does not tell a buyer whether a particular business process can be completed under that buyer's rules. Access to the relevant services is a necessary starting point for connected work, but the useful unit of evaluation is a bounded task with a recognizable completion condition.

Consider the difference between a briefing and an action request. The briefing asks the reader to judge information. An action request asks the reader to authorize a change. Combining those experiences inside a personal agent raises a practical question about how clearly the interface distinguishes them. Buyers should examine that distinction directly rather than treat every agent output as the same kind of result.

The daily morning briefing also creates a natural review point. A team evaluating Duet can use that cadence to compare what was prepared with what was actually authorized. That is a suggested evaluation method, not a claim that the product includes a particular audit feature. The goal is to find whether background assistance reduces work without obscuring responsibility for the next action.

The trade-off is attention at the approval boundary

Requiring approval protects the user's authority over sensitive actions, but reviewing those actions still consumes attention. If an approval request lacks context, the reviewer must reconstruct the task before deciding. If the request is clear, the human can judge the proposed change without repeating the agent's entire investigation. The fact sheet establishes the approval requirement, not how well Duet solves that usability problem.

Research-preview status should therefore shape the evaluation. A buyer can test whether the agent prepares useful work while keeping sensitive execution out of scope until the review process is understood. That approach does not assume the product is unsafe. It separates a documented safeguard from the additional evidence needed to trust it inside a particular business process.

The competitive race does not remove this cost. Meta Muse, OpenAI Dots, and Google's Gemini Agent give buyers alternatives to examine, but their presence is not evidence that any product has resolved the same governance questions. The comparison worth making is between actual task boundaries and actual review experiences, not between product names or broad promises of personal assistance.

What To Do About It

• Check access eligibility: Confirm that the person running the evaluation qualifies for the US subscriber waitlist before planning a business rollout around availability, using the access requirements on the Claude Duet product page and the named model, Opus 5.5.
• Define the sensitive boundary: Write down what a reviewer must see before approving a payment, then compare that requirement with a real product demonstration instead of assuming connector access implies authority across 6,000+ connectors.
• Review background outputs: Compare a prepared briefing with the actions it proposes, recording which conclusions can be accepted and which require further human investigation before the next daily morning briefing in Claude Duet.
• Separate model and permission tests: Evaluate whether a task was prepared correctly independently from whether its execution was authorized, keeping the capability discussion anchored to the announced model, Claude Opus 5.5.

The Bottom Line

Duet's October 21 launch makes proactive background work a concrete product decision rather than an abstract agent discussion. Anthropic's 6,000+ connectors establish breadth, while required approval for sensitive actions establishes a boundary. Neither figure tells a buyer how much judgment an approval will demand. Inspect that handoff before treating a personal agent as a business process.