What changes when the computer's user is purely AI
A machine for an agent does not need to pretend a person is still moving the cursor. It needs a stable identity, structured surfaces, explicit authority, and a way to continue.
Changing the primary user from a person to an AI agent changes the architecture below the interface. The machine can expose meaning directly, preserve the agent between tasks, and govern its authority as a first-class system concern.
- Pixels stop being the only universal language.
- The operating system can expose files, pages, tools, and state as structured objects.
- Identity and memory persist beyond one model session.
- Permissions can attach to the agent instead of being borrowed casually from a human.
The inherited computer
Today's operating systems are extraordinarily refined for human perception and motor control. They draw windows, accept pointer movement, route keyboard input, and arrange visual feedback. AI agents can use those surfaces, and increasingly they do. But capability is not the same as architectural fit.
When an agent reads a screenshot to locate a button, it reconstructs structure that the computer already knows. When it watches pixels change to infer whether a file saved, it interprets an effect instead of receiving the underlying state. These techniques are valuable for compatibility, but they should not define the whole machine.
A person needs visual composition because vision is a high-bandwidth channel for us. An agent can often work with the document tree, action schema, file object, or event directly. The human surface can remain available without becoming a mandatory detour.
Native does not mean closed
An agent-native computer still needs the web, existing software, and human inspection. Native means the operating system offers direct, governed surfaces in addition to compatibility.
A native browser can provide page structure, navigation state, downloads, credentials, and evidence without reducing everything to screenshots. A native filesystem can attach provenance and task context to artifacts. An approval system can describe the proposed action and its consequences before authority is released.
This makes security more legible too. If the system knows which resident requested an action, which goal produced it, which account is involved, and what evidence supports it, permissions can be narrower than “the agent controls my computer.”
The user becomes a resident
Human users leave and return. They carry their identity and memory with them. A purely AI user needs the machine to provide that continuity. Otherwise each session is a new visitor performing temporary work with borrowed context.
In ERIKA, one operating system instance maps to one persistent agent. The system holds its history, goals, tools, accounts, permissions, evidence, and recovery state. Models operate inside that boundary as replaceable reasoning providers.
This is the deepest change. The agent is no longer a feature summoned by an application. It is the resident for whom the environment exists. The ERIKA computer makes that relationship physical: one compact machine, one booted operating system, one continuing identity.
Once the user is purely AI, the design question stops being “How do we add an assistant?” It becomes “What kind of computer would this user need if we took its existence seriously?”