Typical tasks
- Email and calendar briefing
- Connected SaaS work
- Background research and documents
- Follow-ups and proactive suggestions
Evidence-based agent profile · Enterprise & Produktivität
Proactive Claude agent with a cloud computer that prioritizes connected-app work, continues in the background and asks approval for consequential actions; US waitlist preview only.
Direct answer: Proactive Claude agent with a cloud computer that prioritizes connected-app work, continues in the background and asks approval for consequential actions; US waitlist preview only.
This profile summarizes product-specific primary sources, not an independent audit. Missing controls are unknown, not false.
Duet inherits Claude memory, skills and existing connectors and runs on a separate cloud computer. It periodically checks connected services, surfaces work and carries out multi-step tasks. Anthropic describes approvals before sending, signing, paying or deleting.
Useful to invited US professionals with heavy inbox and calendar workloads. Anthropic notes that personal errands, including ticket purchases, are weaker in this early build.
Not yet available on Claude Team or Enterprise. Administrative SSO, audit and compliance features elsewhere in Claude should not be inherited as confirmed Duet capabilities.
Cowork mainly handles knowledge-work requests initiated by users. Duet adds proactive discovery of outstanding work, scheduled checks and follow-up prompts. They share parts of Claude but are separate products with separate access requirements. A Cowork subscription does not itself prove eligibility for Duet.
Duet operates on its own cloud computer and reuses connected Claude services. It produces morning email and calendar briefs and revisits selected sources throughout the day. Multi-step tasks such as forecasts, reports or websites can use many model turns. These runs count against the account's normal Pro or Max limits.
Emails, documents and webpages can contain untrusted instructions. Anthropic explicitly warns about prompt injection and describes approval gates for consequential actions. A vendor's safety design is not equivalent to independent testing. Start with read-only access, audit connector scopes, and check the original evidence for significant outputs.
The research preview is restricted to invited US Pro and Max users; invitations roll out in waves. There is no documented general German, Team or Enterprise availability at this check date. The practical value of proactive action should be measured in controlled workflows before sharing extensive accounts.
The following checklist separates a vendor feature description from the evidence needed before adopting this product. A missing control is not evidence that it does not exist.
Primary sources describe the task scope, but not an independently measured success rate.
US waitlist for Claude Pro and Max; separate iOS/Android app and web, not Team/Enterprise
Actual permission, latency, failure recovery and costs depend on the deployed application.
A pilot should use a read-only connector, a separate calendar and limited test mailbox. Check whether the morning brief prioritizes correctly, how errors can be corrected and how many plan limits periodic checks consume. Record whether approval is visibly required before a draft message is sent. Without a US invitation, first-hand testing is unavailable.
The sources below are original product documentation. Statements about feature availability are scoped to the dated rollout, platform and customer plan. Marketing descriptions of security or compliance do not establish legal compliance for your scenario; review the current vendor terms and request audit artifacts separately.
Product access rules, data processing, contract commitments and action logs must be assessed in the actual tenant. The EU AI Act classification depends on deployment purpose and human oversight, not on the product label alone.
US waitlist for Claude Pro and Max; separate iOS/Android app and web, not Team/Enterprise
Check real entitlement, supported region, usage quotas, connector permissions and cancellation terms against the latest provider documents before purchasing.
These links lead to separate AgentenCode profiles. Similar providers do not share identical trust evidence or access conditions.
What does the manufacturer establish? Vendor documentation establishes only the stated product scope and date. No independent functional or penetration test was conducted.
The documented research preview requires an eligible US Pro or Max account and a waitlist invitation.
No. Duet is a distinct proactive app with background checks and different access rules.
They count against the existing Pro or Max usage limits, including session and weekly limits.
Anthropic describes approval before messages and other consequential actions are sent or executed.
No independent end-to-end Duet security audit was established from these sources.
· New source-scoped profile added. No independent tenant test or security certification has been performed.
Agent history →