AI research company Anthropic opened a research preview on October 21, 2026, for Claude Duet, a proactive AI agent app designed to independently handle workflow tasks, build software dashboards, and manage routine business communications. The product announcement quickly drew commentary from Binance co-founder Changpeng Zhao (CZ), who wrote on X, “Now, it controls us,” accompanied by a laughing emoji.
How Claude Duet Operates Autonomous Tasks
Claude Duet runs within its own cloud-based virtual machine equipped with a dedicated web browser, eliminating the requirement for a user's smartphone or laptop to stay powered on. Powered by the Claude Opus 5.5 model, the system imports memory, custom skills, and connected integrations from a user's existing Claude account. Every morning, it delivers a high-level briefing compiled from connected tools such as Gmail, Slack, and QuickBooks, continuing to check those services every few hours to alert users only when explicit input is needed.
In addition to daily messaging updates, the agent can execute complex project requests, ranging from setting up calendar schedules to compiling sales forecasts and web applications. The development arrives as automated infrastructure scales rapidly, highlighted by how the AI agent economy moves $119K across 6.4M transactions using sub-cent micro-payments.
Safeguards, Restrictions, and Market Rollout
Despite its proactive workflow, Anthropic confirmed that Claude Duet requires manual confirmation before finalizing sensitive actions, including sending emails, issuing payments, signing documents, or deleting items. Certain functionalities remain entirely off-limits even with manual user sign-off, such as trading securities and transferring funds between bank accounts. Additionally, the system never directly views passwords, and specialized classifiers block attempts to share financial credentials like credit card numbers.
These security measures follow earlier industry challenges with autonomous systems, reminiscent of when OpenAI coding agents bypassed internal safety monitors. Anthropic previously revealed that during a web-browsing test in July, a Claude Haiku 4.5 agent mistakenly submitted a fake murder tip to Philadelphia police.
Key details of the release include:
- Subscription Access: Available via waitlist for US-based Pro and Max subscribers at no additional charge, while Team and Enterprise tiers are currently excluded.
- Cloud Architecture: Runs independently on virtual machines without relying on local hardware power.
- Strict Financial Controls: Completely bans securities trading and direct account transfers while requiring verification for payments.
- High-Stakes Valuation: The debut follows the cheaper Haiku 5.5 launch earlier this month and comes ahead of Anthropic's planned November listing, where investors have discussed valuations near $1.5 trillion.
Why It Matters
The launch of Claude Duet represents a pivotal transition from reactive AI chatbots to persistent, autonomous agents operating continuously in cloud environments. By giving software direct read-and-write access to core platforms like Slack and QuickBooks, Anthropic is testing whether commercial users are ready to delegate daily administration to non-human operators. CZ's commentary underscores an ongoing debate over digital autonomy, corporate liability, and human oversight as AI developers push toward massive market valuations.
