Anthropic opened an early research preview of Claude Duet on October 21, 2026, introducing a proactive agent designed to handle tasks in connected apps without waiting for direct user prompts or step-by-step instructions [1]. The new tool runs on cloud systems and brings over user context, memory, and skills from existing Claude accounts [3]. Unlike standard chatbots that sit idle until someone types a question, this software checks linked accounts, prepares morning briefings, and keeps working after people log out of the software [4]. Access opened through a waitlist for paying Pro and Max members in the United States [1].
Claude Duet Core Architecture
Anthropic built the agent around its Claude Opus 5.5 model to give the software enough judgment to separate urgent items from routine background updates. Opus 5.5 guides decisions [1]. The model reads incoming alerts, decides when an alert is necessary, and checks whether a requested job needs human sign-off before proceeding. It operates on a virtual machine equipped with its own cloud browser, which lets the tool interact directly with websites even when direct platform connectors do not exist [3].
Users talk to the agent through voice recordings or typed text within a single ongoing chat thread. Instead of splitting work into scattered chat rooms, the layout keeps context in one main workspace while letting people set up separate topic channels that share the same memory bank across different projects. The system searches past talks so people do not need to repeat basic facts or project details [4]. It handles quick administrative chores like booking calendar events, but it can also take on heavy technical work such as prototyping digital tools, coding web apps, and drafting sales forecasts for growing teams [2]. When multi-step projects finish, deliverables show up in a Make tab for manual review and file export. Finished jobs return as editable files, including research briefs, financial projections, slide decks, and custom dashboards [4].
To track running background jobs, the software includes a Tasks tab that lists every active job and flags pending decisions. According to Claude Duet support notes, the assistant sends live updates to the main conversation thread whenever an assigned task finishes or requires an explicit user decision. Unlike automated scripts that run wild, only human workers hold the authority to mark assignments as complete. The tool can’t close tickets alone [4].
Daily Briefings and Work Tools
Anthropic shaped the daily routine around busy professionals and small business owners who juggle more tasks than they can track alone [2]. Every morning, the agent scans connected inboxes and calendars to assemble a concise briefing that highlights key deadlines, unanswered messages, and emerging project priorities. It then drafts first replies to pressing messages so users can check the wording before sending them out. Marcus Mendes at 9to5Mac said this daily summary helps people start their mornings with clear priorities [4].
Throughout the workday, the agent checks linked services every few hours without needing manual prompts [3]. When an email thread goes quiet or a problem appears, it sends a quick alert to the screen. The platform supports add-ons, including a small business plugin that tunes tools for retail and trade shops [4]. Ben Schoon at 9to5Google reported that Anthropic also launched integrations connecting Claude to Google Docs, Sheets, and Slides with an interactive sidebar interface to help people manage documents. That update arrived alongside broader industry moves tracked by 9to5Google, including Google redesigning Gemini tasks to fold its Spark helper into the main chat application [2].
Because the virtual machine runs in the cloud, work continues after users log off or close their laptops [3]. The remote setup crunches data sets, organizes spreadsheets, and builds slide decks while people attend outside meetings. Mendes said this cloud setup lets workers hand off large multi-step jobs that would otherwise slow down personal laptops [4]. Cloud agents never sleep [3].
Safety Filters and Action Approvals
Giving an automated system access to outside accounts brings real security risks regarding private records [1]. Claude Duet never sees passwords. Anthropic said the agent connects to outside web tools and cloud services without ever seeing or storing user account passwords in plain text. To stop data leaks, developers built a safety classifier and a new system prompt that checks each outgoing message [3]. Jay Peters at The Verge reported that the safety filter blocks prompts that try to expose confidential user records, including credit card numbers [5].
The assistant can’t take irreversible steps without explicit human approval. It pauses and asks for permission before sending an email, signing a legal paper, making a payment, or deleting an existing file [3]. For sales, the tool pairs with Stripe to spot unpaid bills and draft payment reminders. It also links with payment options like Shop Pay and Link to prepare checkout orders that wait for user sign-off [4]. Users retain the final say [1].
These rules keep human choice at the center of automated workflows. Anthropic said Claude Duet builds on years of company research into agent safety and security [1]. Juli Clover at MacRumors said requiring sign-offs on file deletions and electronic payments ensures users retain direct control over software actions. Clover highlighted that like OpenAI Dots and Meta Muse, Claude Duet functions as an agent that reaches out before users ask. By requiring confirmation for risky actions, the setup prevents mistakes in linked accounts [3].
Claude Duet Waitlist and Pricing Limits
Anthropic launched the research preview exclusively for paying Pro and Max subscribers in the United States [1]. Eligible account holders must join a waitlist, with invitations going out in gradual waves. The preview software is packaged as a mobile application for iOS and Android devices, while desktop users can interact with it through the main claude.ai web portal. Invitations roll out in waves [4].
Anthropic does not charge an extra fee to use the preview, but all agent tasks count against existing plan usage limits. The preview is free. Because background tasks run on cloud servers, ongoing account monitoring burns through quota even when a user is not actively typing messages. Marcus Mendes at 9to5Mac noted that these continuous background checks mean active subscribers could run through their standard plan limits much faster than when using standard chat alone [4].
Anthropic technology previously expanded across subscription tiers when Google added Antigravity Claude models for paid Pro and Ultra plans. In this preview, business teams must wait a bit longer for access. Anthropic confirmed that Claude Duet isn’t available on Team or Enterprise tiers yet, though it plans to support commercial organization packages in future rollout phases for larger offices. Enterprise teams must wait [4].
Early Flaws in the Preview
The debut of Claude Duet speeds up competition among big tech firms building proactive assistants. The launch follows OpenAI releasing its own Dots agent a few weeks earlier, alongside rising interest in Meta’s Muse tool [2]. Dots also uses cloud browsers. Both rival systems use cloud computers and internal web browsers to complete tasks without constant user supervision [3].
Anthropic said the preview is an early test that still makes mistakes [1]. The company warned that the agent can misjudge what is urgent and might misread context in long chat threads. Peters at The Verge said it is also not as good at personal errands like booking travel tickets. Tickets can’t be booked yet. The system can struggle when sorting through mixed messages from different people in a shared project channel [5]. Early testing will help Anthropic tune its judgment before opening the tool to more people. Engineers plan to update the software as early users report flaws [1].
Can an automated agent balance continuous autonomy with reliable safety? Early user feedback will shape how Anthropic tunes background permissions and response thresholds in coming updates. Because the agent remains an early research preview, Anthropic plans to tune its urgency thresholds based directly on feedback from initial users. For now, American subscribers on Pro and Max plans can sign up for the online waitlist to test how the agent handles their daily routines in real work environments before broader international releases [1].
- PRESS RELEASE Anthropic. (2026, October 21). We’re opening a research preview of Claude Duet. X. [Article Link]
- ONLINE NEWS Schoon, B. (2026, October 21). Claude Duet is Anthropic’s proactive AI agent answer to Muse and OpenAI’s Dots. 9to5Google. [Article Link]
- ONLINE NEWS Clover, J. (2026, October 21). Anthropic Launches Claude Duet, a Proactive AI Agent. MacRumors. [Article Link]
- ONLINE NEWS Mendes, M. (2026, October 21). Anthropic takes on ChatGPT Dots and Meta Muse with Claude Duet. 9to5Mac. [Article Link]
- ONLINE NEWS Peters, J. (2026, October 21). Anthropic takes on Muse with Claude Duet, its new AI agent. The Verge. [Article Link]
- WEBSITE Anthropic. (2026, October 21). Claude Duet design, safety, and waitlist details. X. [Article Link]
