Anthropic Launches Research Preview of Claude Duet Proactive Agent Application
Anthropic today opened the research preview of Claude Duet, a proactive agent application based on Claude Opus 5.5, which can continue working after the user leaves and proactively confirm actions.
The application inherits the memory, skills, and connected applications of the user's existing Claude account, requiring almost no additional setup, and runs on an independent cloud computer, supporting task commands via text or voice.
Claude Duet can send morning briefings, mark items in the inbox and calendar that need attention, draft replies for user approval, handle unpaid invoice reminders, prepare payroll reviews, or generate briefings, presentations, and dashboards.
The security mechanisms are based on years of work in agent security: agents never directly view passwords, and a new classifier prevents sensitive information leaks. Key operations such as sending messages, making payments, signing documents, or deletions require user approval.
Currently, it is only available to users on the Pro and Max plans in the U.S. through a waitlist, providing standalone iOS, Android applications, and a web version, with usage counted against existing plan limits; Team and Enterprise plans are not yet supported.
Developers and professional users are accessing backend agent capabilities through the waitlist, outsourcing repetitive tasks to continuously running models, while traditional manual operations and tools requiring constant monitoring face replacement pressure. Early user feedback will be used to fine-tune issues of misjudging urgency.
Source: Public Information
ABAB AI Insight
Previously, Anthropic advanced long-cycle agent tasks through Claude Code and computer usage features, with Opus 5.5 leading in agent coding benchmarks such as Terminal-Bench, and introducing adaptive thinking and pre-action classifiers, laying the foundation for embedding judgment capabilities into continuously running products.
Resources have been invested in independent cloud computers, application integration, and security classifiers, allowing the memory and connections of a single account to be directly transformed into backend workflows. The motivation is to convert dialogues that previously required user initiation into proactive push and approval modes, enhancing the usage depth of Pro and Max subscriptions.
This is similar to the path of backend agent products launched by companies like OpenAI, and resembles the evolution of early virtual assistants from passive responses to proactive reminders; the industry is currently in a phase of expanding from conversational models to continuous agents, with product forms shifting from chat windows to standalone applications.
Essentially, this is a technological replacement: long-cycle agent judgment capabilities cover some functions of personal assistants and task management tools, making "interrupt only when attention is needed" the default interaction. The mechanism is that the model continuously monitors in the background and only requests human confirmation at critical points, reducing the cost of sustained attention.
ABAB News · Cognitive Laws
Continuous in the background, confirmation in the foreground
Judgment precedes action, approval remains with humans
The more connections, the deeper the agent