Artificial intelligence assistants are evolving from throwaway chat sessions into durable worker profiles that retain memory, hold permissions, and carry responsibilities across every interaction, according to a technical analysis published this month by The New Stack. The shift centers on a fundamental architectural change: the reusable unit in AI products has migrated from individual conversations to the agent identity sitting above them. Three major platforms—Hermes, Grok Bot, and Claude Tag—have independently adopted this design in recent weeks, signaling that persistent bot identities are becoming the industry standard.

Hermes folded Bot Mode into its desktop application in version 0.20.4, bundling the previously standalone plugin roster by default and introducing a tabbed sidebar that places Sessions alongside Bots. Each profile appears as a labeled bot carrying its own conversation thread, avatar, personality traits, scheduling logic, and pinned language model. xAI launched Grok Bot on August 11 with a sidebar of named Bots, each displaying a title, description, and support for as many as 50 routines per account. Anthropic announced Claude Tag on June 23, provisioning a single organization-level Claude identity that joins human team rosters in Slack, where any channel member can mention it and each thread still creates its own working session. Sessions have not vanished from any of these offerings; they now sit beneath a lasting agent identity that preserves memory and permissions across them.

The report emphasizes that a bot functions as a profile rather than a fresh runtime. In Hermes, which operates under an MIT license with open code, a bot constitutes a Hermes Agent profile—a distinct home directory storing configuration files, cryptographic keys, a SOUL file, memory, sessions, skills, and cron state. Routines are standard cron jobs namespaced to each bot, and handoffs between bots are actual command-line invocations delivered into a canonical bot chat whenever the receiving bot executes next. The standalone plugin described itself as a user interface layered over those primitives, requiring no core patches and no background daemons. Profiles existed well before Bot Mode arrived to give them faces and a sidebar; what has shifted is that the unit of reuse moved from the prompt to the profile itself. Descriptions now carry substantial operational weight: xAI instructs operators to encode a bot's lasting responsibilities, data sources, and boundaries within its description, and bots can delegate work to one another, making the job description resemble a routing table.

The three platforms diverge sharply in how they implement identity architecture, and the report warns that diligence belongs on the identity model rather than the feature roster. Hermes scopes credentials to whichever host runs the application—whether a laptop or remote server—and its own documentation clarifies that profiles are not sandboxes, so the separation addresses state and configuration rather than security. Grok Bot takes the reverse approach, with every bot sharing one account-scoped computer that includes files, browser sessions, and application logins; xAI states that these assets survive even after deleting the bot that created them. Claude Tag provisions an organization service identity with channel-scoped access to tools. A named bot with a job title reads like a coworker, but what actually authenticates into a production system is a profile directory, a shared machine, or a service account, depending on the vendor. The counterintuitive reality is that persistence here is mostly a configuration matter: nothing in a profile directory, a shared computer, or a service account guarantees an agent remains coherent across weeks of accumulated work, and none of the three vendors has published evidence that any of them do. Teams assessing any of the three should begin with the identity architecture, because while the roster is simple to replicate, the underlying identity model is not. Organizations that mistake interface polish for security boundaries may discover too late that their AI colleague's credentials live in places they never intended.