# Claude Duet Makes Attention the Agent’s Job - Date: 22 Oct 2026 (2026-10-22T16:08:59.000Z) - Summary: Anthropic’s Claude Duet preview delegates monitoring and background work while retaining approval gates for consequential actions. Its central test is whether proactive assistance saves more attention than supervision consumes, within shared usage limits. - Tags: `digest`, `ai-discourse`, `claude-duet`, `anthropic`, `ai-agents`, `human-oversight`, `ai-workflows` ## Sources 1. [Anthropic - Introducing Claude Duet](https://claude.com/resources/articles/claude-duet) (website) 2. [TechCrunch - Claude Duet app is Anthropic’s bet on agents](https://techcrunch.com/2026/10/22/claude-duet-app-is-anthropics-bet-on-agents/) (website) ## Executive Summary Delegating work becomes a different proposition when the software also decides what deserves your attention. Anthropic’s [Claude Duet research preview](https://claude.com/resources/articles/claude-duet) packages that proposition into a persistent assistant for professionals and small businesses: it checks connected applications, prepares work in the background, and brings decisions back for approval. The consequential change is the interaction model, not a demonstrated breakthrough in reliability. Duet asks users to trust its judgment about when to interrupt them, while retaining approval gates for consequential actions. Its background activity also consumes the same usage allowance as ordinary Claude work. Whether this saves attention rather than creating another inbox remains unproven. ## What Happened Announced October 21, Duet is initially a waitlisted research preview for US Pro and Max subscribers. Anthropic describes standalone iOS and Android apps and access through claude.ai, built around Opus 5.5. It carries over a user’s Claude memory, skills, and connected applications into one ongoing conversation. The operating pattern is concrete. Duet sends a morning briefing, checks connected accounts every few hours, and messages when it judges that something needs attention. It can also take assigned tasks—from scheduling meetings to building dashboards or web applications—and continue working through its own cloud-hosted virtual machine and browser after the user’s devices are switched off. [TechCrunch’s coverage](https://techcrunch.com/2026/10/22/claude-duet-app-is-anthropics-bet-on-agents/) places the launch alongside competing personal and business agents. But the useful distinction in Anthropic’s own account is its emphasis on professional coordination rather than general errands. The company acknowledges that booking tickets is not a strength and that users cannot yet see its browser activity. The examples are still vendor-selected testimonials: a property showing flagged for the wrong day, preparation for a client call, and invoice work across restaurant locations. They illustrate the intended use, not measured evidence of dependable performance. ## Why It Matters **Proactivity makes prioritization part of the product.** A chatbot can answer the question it receives; Duet must additionally decide which situation merits a question or notification. Anthropic explicitly says it is still tuning that judgment and that the preview may misread a thread or miscalibrate urgency. A useful briefing and a distracting stream of plausible alerts can come from similar underlying capabilities. **Approval is the proposed boundary, not the absence of autonomy.** Anthropic says Duet can read, draft, and prepare independently, but must show its proposed action and wait before sending a message, making a payment, signing a document, or deleting anything. Securities trading and transfers between accounts remain prohibited even with approval. The design separates preparing work from committing it; it does not eliminate the burden of checking what is prepared. **Information handling needs its own safeguards.** Permission to send a message does not, by itself, establish that every detail in it should be shared. Anthropic describes a dedicated classifier and safety prompt intended to block disclosure of credit-card numbers and government identification, while requiring explicit instructions for other sensitive information. Those are design claims, not proof that leakage or prompt injection has been solved. **Background work has an opportunity cost.** Checks count against the same plan limits as other Claude activity. Anthropic says accounts are checked every few hours partly because continuous checking would exhaust those limits faster. “Always available” therefore does not mean continuous observation or unlimited execution. ## The Bigger Story This sharpens a recurring distinction in agent discourse: producing plausible work is different from establishing that the work is correct and worth doing. With a proactive assistant, the evaluation also has to include missed urgent events, unnecessary interruptions, and the effort required to approve actions safely. Duet’s launch makes those questions commercially concrete. Its promise is not merely fewer manual steps, but less coordination overhead across applications. The unresolved test is whether the attention it returns exceeds the attention required to supervise it. The preview establishes a product direction; it does not yet settle that trade-off.