Anthropic has opened the research preview for Claude Duet — a standalone application where Claude functions as an always-on cloud agent: it independently sends morning summaries, monitors email and calendar, and continues executing instructions in the background between user sessions. At launch, the service is available to Claude subscribers in the US via a waitlist and is directly positioned against OpenAI's ChatGPT Dots and Meta Muse. Anthropic is moving proactive assistants from niche experiments to a mass consumer product.

image
image

What happened

The announcement was made on October 21, 2026. Duet is designed as a standalone application, not just another tab in a chat: the agent is allocated its own cloud computer, where instructions continue to be executed even after the user has closed the session. The agent operates on the existing Claude Opus 5.5 model and relies on the state of the Claude account — memory, skills, and connectors. A total of over 6,000 integrations are supported, including Gmail, Google Calendar, Stripe, Notion, and Slack. Before sending messages, making payments, signing documents, and deleting data, Duet is required to request explicit confirmation. The app is available on iOS, Android, and on the web at claude.ai. The research preview is available to Pro subscribers with a $20/month plan and Max subscribers in the US; sign-up is via a waitlist on the product page; Team and Enterprise plans are not yet supported.

Context

Duet is not a new model, but a new form of working with an existing one: the engineering novelty is focused on the agent system on top of Claude Opus 5.5, specifically in persistent memory, a set of skills, a dedicated computing environment, and autonomy rules. The agent's state is tied to the account, not to a separate dialogue, so Claude accumulates context from session to session, rather than starting each chat from a blank slate. The event fits into a general market shift: three major labs — Anthropic, OpenAI, and Meta — are releasing always-on agents almost simultaneously, and competition is moving from chat interfaces to the reliability of background task execution. In this field, not only the model's intelligence is important, but also the autonomy contract: which steps the agent takes on its own, and which it leaves for human confirmation. Notably, Anthropic is introducing such a construct not as a corporate pilot, but immediately as a consumer subscription, although team plans have not yet been scaled.

Why this matters for the industry

The announcement provides a public example for builders of agent products to follow: background execution in a dedicated environment, account state instead of chat state, and explicit confirmations for sensitive actions — this UX contract can already be built into their own architectures today. The subject of competition is also changing: the decisive factors become the breadth of integrations, the quality of memory between sessions, and the predictability of the agent's rights, and the industry's key question shifts from which model is smarter to which system is more reliable without supervision. The third effect is economic: background work is deducted from subscription limits, meaning an agent with constant activity significantly increases inference consumption per user, and compute costs become a visible part of the product, not an internal metric. A window opens for startups for vertical always-on solutions in niches that horizontal agents do not yet cover. A limiting factor is the closed nature of the preview: the API is not mentioned in the materials, so Duet cannot yet be integrated into production or workflows.

Why this matters for users

Access today requires meeting three conditions: a Claude Pro or Max subscription, being in the US, and a spot in the waitlist — the application is submitted on the claude.com/product/claude-duet page. No separate payment is required: the app is included in the existing plan, however, background work consumes limits faster than a regular dialogue, so the volume of agent load will have to be dosed. The confirmation mechanics work in the user's favor: sensitive actions will not occur without explicit human consent, which reduces the risk of irreversible steps by the agent. The practical meaning for the reader is to compare the approaches of different labs to proactivity in real work and understand whose implementation of the always-on mode is more convenient for everyday processes. Users outside the US remain to watch for new waves of invitations.

What is currently unknown / limitations

The announcement materials do not contain reproducible measurements: no benchmarks, no eval reports, no methodology, so the claimed proactivity and reliability of background tasks remain marketing claims without independent verification — the preview is closed and no public tests exist. Exact limits for background load and latency characteristics of the dedicated cloud computer are not disclosed, nor is how memory quality behaves during long-term agent operation. The full list of actions requiring confirmation has not been published. The timeline for leaving the waitlist and plans for expanding geography and plans are unknown.

Sources

Author

Look at AI, editorial team