Anthropic opens Claude Duet research preview: a standalone proactive-agent app on Opus 5.5 that briefs you daily, works in its own cloud computer and asks before it sends, pays or signs

Why it matters

【Key points】

On Oct 21 Anthropic opened a research preview of Claude Duet, a standalone app where Claude acts as a proactive agent: it flags what needs you, takes tasks, keeps working while you are away and asks before acting. US Pro and Max users only, via waitlist.

1. iOS/Android app plus claude.ai on desktop, separate from the main Claude app. Built on Claude Opus 5.5 and inherits memory, skills and connected apps.

2. Sends a morning brief and checks connected accounts every few hours (to save plan limits). Tasks run on a cloud VM with its own browser.

3. Sending messages, payments, signing or deleting require explicit approval; trading securities or moving money between accounts is blocked even with approval. A new classifier blocks sharing card numbers and government IDs; credentials follow the Managed Agents design.

4. Uses 6,000+ Marketplace connectors (Gmail, Slack, QuickBooks, Stripe, Gusto, Shopify) and can buy via Stripe Link virtual cards.

5. Shares plan limits, background checks included. No Team/Enterprise yet.

【Verifiable】

https://claude.com/resources/articles/claude-duet

https://claude.com/product/claude-duet

https://x.com/claudeai/status/2112984664256241837

【Our read】

1. Background: assistants are going proactive — ChatGPT Pulse (2025) briefs, Notion agents run 24h+ on their own computers, ServiceNow/NVIDIA just launched Arc. Duet is Anthropic's first consumer app bundling alerts, a cloud computer and delegated action.

2. Comparison: Pulse informs; Duet executes small-business admin like invoices, payroll and contracts. Its guardrails — per-action approval, hard bans on financial transfers, tiered data blocking — are more conservative than most agents.

3. Impact: judgment and cost are the real tests; Anthropic admits it may misjudge urgency, and polling every few hours trades capability for usage. Success would push subscriptions from chat toward delegated work.

4. Caveat: US-only waitlist; curated testimonials, no independent tests yet.