When will Anthropic release a personal AI agent?

If an Anthropic product has implemented a general-purpose personal AI agent that remembers the user and starts multi-step tasks across third-party apps on its own before Jan 1, 2027, then the market resolves to Yes. The function must be live and processing user interactions, not explicitly labeled as "beta," "test," "experimental," "preview," or "early access" or similar, and officially acknowledged by the company through announcements, documentation, or help pages. Implementation includes direct integration into the product's interface, automatic application to user accounts, user-initiated activation through official controls, API implementation, or policy changes that operationally enforce the function. Announcements without deployment, employee-only testing, technical documentation without implementation, and third-party additions without authorization do not qualify. The implementation must occur in the United States unless otherwise specified. If it meets these conditions, it counts from the first date it is live for users outside the company, including in a gradual rollout or a waitlist, invitation-only, or paid-tier release. A general-purpose personal AI agent acts for an individual user on that user's own goals, is not limited to a single category of task or app, such as only code, only documents, or only shopping, and (1) retains user-specific information, such as preferences, relationships, goals, or ongoing activities, across separate interactions, and (2) as a built-in capability, on its own initiative and outside any conversation or task the user has started, begins new multi-step actions across third-party apps, websites, or services to advance the user's ongoing goals, based on what it has retained. The agent must itself take at least one such action before asking the user anything, and may then pause for approval before a step such as sending a message or making a purchase. An action means changing something in a third-party app, website, or service on the user's behalf, such as sending, posting, booking, purchasing, submitting a form, or creating, editing, or deleting an item; reading, searching, or summarizing does not count. An agent focused on work qualifies if it acts across email, calendar, messaging, and the web for an individual user. The following do not qualify: carrying out a task the user started, however long-running; running tasks on a user-defined schedule, event trigger, or monitor, or under a standing instruction that tells the agent what to do or which app, source, or workflow to watch; a briefing or notification on its own; a proposed task or suggestion that does nothing until the user accepts it; an offer made in reply to the user's message; memory alone; an agent configured by an organization to run shared team workflows; and a developer platform, SDK, or API for building agents, unless the company itself offers through it an agent that meets (1) and (2) for end users. A goal or preference the user has shared, from which the agent itself decides what new task to start, is not a standing instruction, and neither is turning the capability on, connecting apps, or granting permissions.