Anthropic has launched Claude Duet as a research preview — a proactive AI agent built on Opus 5.5 that operates more like a remote employee than a chatbot. Instead of waiting for you to open a conversation, Duet initiates contact: morning briefings, inbox triage, flagged calendar items, drafted replies awaiting approval. The interaction model is messaging, not prompting. One persistent thread, voice or text, no setup required. The product surface is deliberately broad. Duet handles scheduling meetings, building dashboards and decks, finding unpaid invoices, drafting contract reminders, preparing payroll for review, and assembling financial models. It runs on its own cloud computer, uses your connected apps, and returns finished artifacts — documents, spreadsheets, presentations — for your sign-off. This is Anthropic's play for the agentic middle ground: not a copilot riding alongside you, but a delegate you hand work to and check back on. The trust architecture is the most consequential design choice. Duet asks before sending messages or making payments. Approvals, logins, and safety checks run in the user's own cloud environment, keeping sensitive credentials out of Anthropic's infrastructure. Users can see what Duet has done and is doing, and can steer it. This is a genuine structural commitment to human-in-the-loop control — not a disclaimer, but an operational constraint baked into the product. The business model, however, is where extraction pressure lives. Duet shares your existing Pro or Max plan usage limits, and it consumes them differently than a typical conversation. Background checks — the morning briefs, the inbox scans — count against your quota. Larger tasks eat more model turns. Anthropic is effectively metering the ambient intelligence layer on the same budget as your direct conversations, which means heavy Duet users will hit ceilings faster and face pressure to upgrade. Team and Enterprise plans are excluded from the preview entirely. The competitive positioning is unmistakable. Google's Gemini and OpenAI's ChatGPT have been layering agent capabilities into existing chat interfaces. Anthropic is instead shipping a standalone app (iOS, Android, web at claude.ai) with a fundamentally different interaction paradigm: the AI comes to you. This is a bet that the winning UX for AI agents is asynchronous delegation, not synchronous collaboration. If correct, it reshapes how knowledge workers allocate attention. The risks are real and honestly disclosed. Duet may misread threads, miscalibrate urgency, or miss items between check-ins. These are not edge cases — they are the core failure modes of any system that decides what deserves your attention. The product is selecting and filtering your information environment, and errors in that selection have compounding consequences. An assistant that misses an urgent invoice or over-escalates a routine email trains you to distrust it, collapsing the delegation loop. The 20-year trajectory here splits sharply. If Anthropic maintains the human-approval architecture and transparent activity logs, this could become a genuinely generative tool that returns time to knowledge workers. If the metering model tightens and the approval gates relax under competitive pressure, it becomes a subscription funnel that concentrates workflow dependency on a single vendor while extracting escalating fees for ambient compute.
