Anthropic launches Claude Duet: an always-on agent that keeps working after you power down — 6,000 connectors, approval gates, and a direct challenge to Muse and Dots

The AI assistant is evolving from "answer me" to "do it for me": on Wednesday Anthropic launched Claude Duet, an agent app that runs continuously in the background, proactively identifies needs and executes tasks — powered by the flagship Claude Opus 5.5, able to call over 6,000 tool connectors, and able to keep working after you shut your devices^[1]. TechCrunch's framing names the battlefield: this is Anthropic's answer to Meta Muse and OpenAI Dots — the proactive-agent three-way war has formally begun^[2].

The form factor: from chat box to "digital employee"

The fundamental difference from conversational assistants is proactivity. Per Anthropic's announcement, Claude Duet targets busy professionals and small-business owners: every morning it rounds up important items from connected apps, keeps checking information through the day, and proactively alerts when something needs the user; users can also assign tasks directly — from administrative work like scheduling to building dashboards with multiple tools, producing financial or sales forecasts, and developing prototypes and web apps^[1]. The interaction model changes accordingly: no need to stay in the chat window issuing step-by-step instructions; the agent advances work in the background and notifies you when done or when review is needed. Duet inherits the memory, skills and connected apps of the Claude account, ships as standalone iOS/Android apps and on the web; with a dedicated virtual machine and browser environment, it can keep executing after your computer or phone is off — the technical foundation of the "always-on" positioning^[1].

Safety design: approval gates and a "cannot-execute" list

Anthropic's safety arrangements are careful: critical steps — payments, sending messages, signing documents, deleting information — display the proposed action and wait for approval; a dedicated mechanism guards against improper disclosure of sensitive information; and some high-risk operations cannot execute even with user approval^[1]. The design targets the agent trust problem at its core — when AI can move your money and your email, an auditable approval chain matters more to the product's fate than a clever model. The limitations are stated honestly: Anthropic admits Duet still falls short in judging urgency and handling some personal affairs, and the product is in research preview, open only via waitlist to US Claude Pro and Max subscribers — the limited scope means it has not yet entered full competition^[1].

Three-way war: positioning differences among Duet, Muse and Dots

Side by side, the positioning is distinct. Meta's Muse takes the consumer-scale route: over 10 million US-Canada users, shopping the hottest scenario, snowballing on Meta's social graph and commerce loop; OpenAI's Dots rides ChatGPT's installed base and developer ecosystem; Anthropic's Duet bets on "professional productivity" — 6,000 connectors spanning Gmail and Slack to QuickBooks and Stripe, aimed at heavy workflow users who will pay to save time^[1][2]. Three routes, three monetization hypotheses: transaction take-rates (Muse), subscription upsell (ChatGPT/Dots), productivity-tool fees (Duet). Whichever works first will define the revenue template of the agent era.

What to watch

First, Duet's pace from research preview to general availability and its pricing — standalone pricing beyond Pro/Max will reveal Anthropic's revenue expectations for agents. Second, the real-world economics of "works after shutdown" — no one has yet validated the unit economics of a resident background agent. Third, the friction of the approval gates — too much and users churn, too little and risk runs loose; that balance decides whether agents enter true production environments. Fourth, a head-on collision in shopping — Meta has proven transaction value; if Duet moves into procurement and expense workflows, the war burns onto the same battlefield^[1][2].

The compute ledger: the cost reef of resident agents

Duet's "keeps working after shutdown" is magic on the experience side and a cliff on the cost side. An agent with a dedicated virtual machine and browser environment polling information sources around the clock burns orders of magnitude more compute than an on-demand chat — Meta admitted on its call that Muse "consumes extreme compute," which is why it can't launch globally yet: the same ledger. That explains why Duet is gated to a small pilot of Pro and Max subscribers: subscription fees are the most direct compute hedge, and the waitlist is both a cost-control gate and a probe of willingness to pay. The next leg of the agent race may turn not on whose model is smarter, but on who first drives "cost per task" below the commercial viability line — Anthropic promising tenfold cost-efficiency gains next year, Meta the same, both placing the identical bet on the same day.