Anthropic's Claude Duet brings always-on AI agents to mobile devices

Anthropic has introduced Claude Duet, an always-on, proactive AI agent currently available in research preview for U.S. subscribers on Pro, Max, and Ultra plans. Unlike standard chatbots that wait for user prompts, Duet operates on its own virtual machine with a browser, delivering morning briefs, checking connected apps every few hours, and processing background tasks.

What happened

According to the Zapier Blog, Claude Duet functions as a separate iOS, Android, and web app that integrates with a user's existing Claude account. During onboarding, users can connect first-party apps like Gmail, Slack, Calendar, and Google Drive with a single tap. Through a dedicated connection with Zapier, Duet gains access to thousands of additional applications.

The agent features three core sections: Chat for text or voice interaction, Tasks for background work, and Make for building projects using templates like email triage and daily to-do lists. While Duet can autonomously read, draft, and prepare work, Anthropic states that the agent always requests user approval before sending messages, making payments, signing documents, or deleting files. Furthermore, it employs OAuth to protect credentials and screens for prompt injection and sensitive data leaks.

Early testing highlights both promise and practical constraints. Testers noted accurate proactive summaries—such as flagging and explaining Google security alerts overnight—but also found that background checks and tasks deplete Claude session limits quickly. To mitigate this token consumption, users can prompt Duet to set up deterministic workflows on platforms like Zapier for repetitive tasks, reserving agent reasoning for jobs requiring genuine judgment.

Why it matters for operations teams

For small team leads and operations managers, Claude Duet represents a shift from reactive prompt engineering to autonomous orchestration. However, the early reality check around usage limits proves that running a heavy reasoning agent for routine, predictable workflows is an expensive anti-pattern. If a task executes identically every day—such as extracting invoice details and updating a spreadsheet—relying on an always-on AI agent will burn through token quotas in days.

The sustainable path forward is hybrid automation. Use proactive agents like Duet for exploratory triage, exception handling, and unstructured reasoning where context matters. For repetitive, high-volume mechanics, offload the execution to deterministic workflows. Just as video assistant referee (VAR) technology in sports provides evidence while humans retain final authority—as discussed in recent UiPath analyses on agentic testing—operations teams must keep humans accountable for final business decisions while letting automation handle the predictable execution layer.

What to do this week

1. Audit your team's current recurring workflows to identify which tasks truly require AI judgment versus those that run on fixed, deterministic logic.
2. Review your team's active Claude or AI plan usage limits to understand how background polling and token consumption will impact daily operations if you adopt proactive agent previews.
3. Test hybrid execution by setting up a simple multi-step workflow where an AI agent handles unstructured input extraction, but passes the structured payload to a traditional automation tool for execution.

Sources

• What is Claude Duet? — Zapier Blog
• What VAR can teach us about agentic testing — UiPath Blog