Summary
Anthropic has opened a US waitlist for Claude Duet, a research-preview agent for Pro and Max subscribers. It can work in a cloud virtual machine with a browser and connected apps while a user is offline, but it shares existing plan usage limits and pauses for approval before sensitive actions.
What happened
Duet is a standalone app for iOS, Android, and claude.ai, built around Claude Opus 5.5 and carrying over users’ Claude memory, skills, and connected apps. Anthropic says it can use more than 6,000 Marketplace connectors and plugins, and describes tasks such as preparing briefs, drafting replies, and handling business paperwork. It checks connected accounts every few hours rather than continuously. Messages, payments, signatures, and deletions require approval; securities trading and transfers between accounts are unavailable even with approval. The preview is included in Pro and Max, but background account checks and larger tasks count against their existing usage limits. Anthropic has not published independent reliability, security-effectiveness, or task-completion measurements.
What the archive adds
The archive adds a useful precedent: its Browser Use fragment describes browser agents developing beyond an open-source library into a broader ecosystem spanning managed cloud, stealth, and MCP. That points to a shift from isolated browser automation toward packaged infrastructure. The dossier also says Anthropic’s safe-agent framework frames agents as systems pursuing goals with tools and limited human input; this is context, not proof that Duet performs reliably in production.
Trend signal
The product marks a move from answering prompts toward delegated work that persists while a person is away. Its broad connector set makes access look abundant, but the actual operating model is deliberately bounded: periodic checks, approval gates, and shared subscription capacity. The likely commercial bet is that users will accept less immediate control in exchange for getting back completed work. The preview does not yet show whether that exchange works at useful scale.
Why now
The shift toward agents pursuing goals with tools is consistent with Anthropic’s safe-agent framework, according to dossier finding d1. Duet’s own shared limits and periodic account checks make persistent work a product and capacity trade-off, not just a model capability. According to dossier finding d2, Claude Cowork already supports cloud sessions and scheduled tasks, suggesting Anthropic is extending an existing persistence direction. The announcement’s approval gates reflect the practical need to retain human control over consequential account actions.
Connections to other events
The consequential constraint is the shared quota. A task that keeps checking accounts or requires substantial work consumes the same allowance the subscriber uses elsewhere in Claude; users therefore have reason to care about whether each delegation saves meaningful effort. Approval gates also mean Duet is not simply an unattended operator: it can prepare actions, but the user remains in the loop for consequential steps. Our read is that adoption will turn less on the headline number of connectors than on whether routine tasks finish accurately without excessive usage or approval friction.
Why it matters
For builders and buyers, Duet is a signal to evaluate agents as workflows with bounded permissions and measurable completion, not as a list of integrations. Buyers should identify recurring tasks where a prepared draft or report has standalone value even if a human must approve the final action. Anthropic, meanwhile, can use the preview to test whether persistent work fits a subscription model before offering evidence that the limits and safeguards are adequate for heavier use.
What happens next
- Anthropic will publish or surface at least one concrete Duet preview signal beyond connector count, such as task outcomes or usage behavior, as it evaluates the product. (6 months, 55%). Watch for: A Duet product update or preview report includes a task-completion, usage-limit, or user-adoption measure.
- The most visible near-term Duet use cases will remain preparation and drafting rather than autonomous execution of consequential transactions. (6 months, 70%). Watch for: Product materials or release notes emphasize briefs, drafts, and approval-mediated actions while retaining restrictions on high-impact operations.
- Anthropic will broaden Duet access beyond its initial US Pro and Max waitlist if preview demand and operational performance support it. (12 months, 45%). Watch for: An official announcement adds another region or subscription tier to Duet availability.
Limits and uncertainty
The announcement gives no task-completion rates, independent security assessment, or practical account of how much work a subscriber can run before reaching plan limits. The every-few-hours check cadence may not suit time-sensitive workflows, while mandatory approvals may erase some of the convenience. This take could be wrong if users value occasional asynchronous preparation enough that quota consumption and intervention are minor concerns.
Skill to build now
Practice writing bounded agent briefs. First exercise: choose one repetitive, low-risk task; specify the exact inputs, permitted tools, output format, prohibited actions, approval points, and a short human verification checklist.
Reader takeaway
This week, pick one recurring, low-risk task—such as preparing an invoice follow-up draft—and write a bounded brief: permitted data, allowed outputs, actions requiring approval, and a checklist for checking the result. If you test Duet when access becomes available, track time saved, corrections needed, approvals requested, and usage consumed; connector breadth alone is not a success metric.
Open question
Can Anthropic make persistent work valuable within shared Claude usage limits while keeping enough human approval to preserve trust without turning delegation into another queue of chores?