Anthropic opens research-preview waitlist for Claude Duet personal agent

• Duet is designed to spot work and start it before being asked, according to the waitlist announcement.
• A separate mobile and web app organizes channels, tasks, documents and slides, as described in a two-week tester’s account.
• Users cannot take over its cloud browser yet, a limitation an early comparison flags.

A YouTube performance dashboard began with a channel URL and a plain request for insights. Anthropic says a brand strategist received a client-call brief and a warning about a flight conflict without asking.

Duet runs on Opus 5.5, with three operating mechanics described in Anthropic’s launch post:

• Its own cloud computer: A virtual machine and browser keep working even when the user’s laptop or phone is off.
• Existing account context: Memory, skills and connected apps carry over from Claude.
• Periodic checks: Duet checks connected accounts every few hours and surfaces items needing attention. Users can change what it reports and how often.

Dan Shipper, whose team tested Duet for two weeks, described a standalone iOS, Android and web app with:

• Channels for different areas of work and life.
• A to-do list that keeps work and its status visible.
• In-app editing for documents, slides and other deliverables.

The visible work queue is the best design choice here.

Peter Yang, an early-access tester, singled out Artifacts as Duet’s possible “secret sauce” in his early-access review. His full walkthrough demonstrates five concrete workflows:

1. An HTML morning brief. Typing / exposes skills, including a built-in morning skill that renders the day’s priorities as a styled artifact.
2. A YouTube analytics dashboard. A request for performance insights plus a channel URL produced analysis of a year’s public videos, including publish-date charts and solo-versus-interview comparisons.
3. A sponsorship dashboard. It organized booked revenue, sponsor concentration and conversations needing follow-up. The demonstration used mock financial data; Yang also described scheduling recurring dashboard updates.
4. A writing-style guide. A skill read 33 sent emails, 89 Slack messages, 47 DMs and 11 documents, then saved a guide for future drafting.
5. An AI Explorer Club for his child. The artifact included an animated robot, an adventure path and chatbot lessons. Yang found the first version too text-heavy.

Tasks appear as subthreads beneath the main conversation, with active, upcoming, scheduled and resolved views. Yang’s account had a default check-in scheduled every three hours between 9 a.m. and 10 p.m., which he felt was too frequent.

Yang’s attempt to book an Uber stalled at a human-verification check. He could not clear it manually because Duet offered no way to view or interact with its cloud browser.

Shipper reported three other limitations in his two-week test:

• Voice: Transcription only, without a spoken back-and-forth conversation.
• Other Claude sessions: Duet could see sessions on his computers but could not direct them.
• Email drafts: He preferred Dot’s drafts.

Anthropic describes three boundaries in its safety explanation:

• Approval required: Before sending messages, making payments, signing documents or deleting anything, Duet shows the proposed action and waits.
• Off-limits even with approval: Trading securities and moving money between accounts.
• Sensitive-data controls: A dedicated classifier and safety prompt are designed to reject requests to share credit-card numbers or government identification. Sending health or personal-finance information requires explicit instructions.

Access is through a research-preview waitlist for Pro and Max subscribers in the US, the geographic restriction noted in a rollout report. Team and Enterprise plans are excluded, according to Anthropic’s getting-started guide.

Duet shares the same usage limits as ordinary Claude conversations. Background checks count toward those limits, while larger jobs such as decks and financial models require more model turns and background work.

Yang reported little usage on his Max plan in a follow-up, attributing that to what he considered Opus 5.5’s greater efficiency.