What is a personal AI agent?

A personal AI agent is a subscription assistant that lives inside the messaging apps you already use and acts on your behalf over time: it reads email, checks the calendar, drafts replies, books and orders, remembers preferences and habits, and messages you first when it finds something useful. The shape is the same in every case: an agent hosted by the vendor on a cloud computer reserved for one user, with memory across sessions and proactive behavior. Several vendors arrived at it in 2026: Google with Gemini Spark, xAI with Grok Bot, Meta with Muse, OpenAI with dots, and a smaller product, Instinct. Do not confuse it with Muse Glimmer, the open weights model from the same Meta family: it is a model, not an end-user product. The real change from the past is the packaging: what used to be an experiment to install and maintain yourself becomes a mass product where the vendor runs execution, isolation and updates.

See also: OpenClaw · A self-hosted personal AI agent that went viral in 2026, then was hit by a critical vulnerability and hundreds of malicious skills in its registry.

From self-hosted experiment to hosted service

The direct predecessor is the self-hosted open source personal agent you install yourself, whose best known case is OpenClaw: the same idea of an assistant that reads mail and takes action, but with setup, API keys (or a linked subscription) and security left to the user. The hosted wave flips the model: the vendor provisions a separate cloud environment per user, and the agent does not start on its own: you switch it on, it stays under your direction, and it checks with you before the actions that count. The interface stays conversational, inside dedicated apps or the chat people already open every day, sold as a subscription. The execution side, wake cycles and continuity over time, is what persistent agents describe, except here the vendor carries it behind a finished product.

It remains a commercial product with moving boundaries: availability, features and integrations vary by country and plan, so verify them when you choose rather than assuming they hold. For a company that is the sensitive point: an assistant touching mailbox, calendar and payments inherits the permissions of whoever uses it, and every autonomous action must leave a trace that can be reconstructed. Before connecting it to company systems, set up separate work and private roles, keep an inventory of active integrations, and keep a log of what it did without being asked.

An enterprise example

A services SMB gives each salesperson a personal agent connected to company mail and calendar, with one simple rule: reading and drafting yes, sending and paying never without review. The agent triages the morning inbox, proposes replies, holds slots for site visits and drafts quotes. When it finds a payment due it messages the owner instead of running it. After six months the accumulated memory, client preferences, recurring wording, deal history, is worth more than the subscription itself: switching vendors would mean starting over, and that is where lock-in happens.

Why it matters for decision makers

Three decisions belong to whoever signs. First, permissions: read access before write access, never the owner's personal credentials, and payments always with a second confirmation. Second, governance of proactive behavior: what it may do unprompted, at which hours, with what duty to summarize, because an assistant writing to clients and suppliers commits the company. Third, memory: where it lives, who can export it, what happens at cancellation. A personal agent wired into company systems without these answers is shadow AI under a friendly name.