Sovereign AI for Individuals, Not Nations
Search for "sovereign AI" and you will find nations. Countries building domestic model training capacity, governments procuring GPU clusters they control outright, regulators drawing borders around where citizen data may be processed. The word means something specific there: a state that does not need another state's permission to compute.
That is a real and important use of the term. It is also not the only scale at which the idea makes sense.
The same argument, one person wide
Strip the national framing away and the argument underneath is simple. A capability you depend on, which someone else can revoke, price, or redefine without asking you, is not really yours. Nations noticed this about model infrastructure. The same thing is true of the AI agent that reads your notes, remembers your projects, and increasingly acts on your behalf.
A sovereign AI agent is an AI agent whose identity, memory, skills, workspace, and runtime are owned and governed by its user rather than by a model provider or platform. Its durable state lives in files you can read, edit, version, and move. Its actions run inside permission boundaries you define. And replacing the underlying model does not reset the agent that knows your work.
Nations want sovereign AI so a foreign vendor cannot switch off their compute. You want a sovereign agent so a vendor cannot switch off your memory. It is the same objection at a different scale.
What you actually rent today
Most people's "AI assistant" is an account. Consider what that means concretely:
- The memory of your past conversations lives in a database you cannot read, export in full, or correct.
- The assistant's personality and instructions are configured through a form whose behaviour changes when the vendor updates the model beneath it.
- Its ability to act on anything — files, email, calendars — exists only through integrations the vendor chose to build and can retire.
- If you stop paying, or the vendor changes terms, or the product is discontinued, the accumulated context is gone. Not degraded. Gone.
None of that is scandalous. It is the normal shape of a hosted product, and hosted products are often the right trade. But it is worth naming plainly, because the thing being accumulated in that account — a model of how you work — is unusually expensive to rebuild.
Sovereignty is not the same as self-hosting
The most common misreading is that sovereignty means running something on your own hardware. Local execution is part of it, but it is neither sufficient nor the whole point.
You can run a model locally and still have a non-sovereign agent: if its memory is an opaque binary blob, if its identity is welded to the specific model weights you happen to be running, if there is no way to inspect or constrain what it does with your filesystem, then "local" bought you privacy but not ownership.
Conversely, an agent can call a hosted frontier model over an API and remain sovereign, provided the parts that constitute the agent — memory, skills, configuration, permissions, workspace — stay on your side of the line. The model is an engine. Engines are replaceable. That is the point.
Agent sovereignty has four conditions:
- Sovereign mind — the model is an engine, not the self.
- Sovereign authority — autonomy operates under rules you write.
- Sovereign vessel — the runtime is a domain you control.
- Sovereign continuity — the agent outlives any platform it ran on.
Each is spelled out, with a five-question test for checking whether any given agent actually meets them, on what makes an agent sovereign.
Why the distinction matters now
Agents are crossing from answering to acting. An assistant that drafts a paragraph and an agent that files the expense report, pushes the commit, or sends the message are different kinds of thing to hand your life to. The second kind accumulates a great deal of context about you, and exercises a great deal of authority on your behalf.
The moment to decide who holds that context and that authority is before it is substantial enough to be painful to lose.
Nations arrived at this conclusion about compute. The version for individuals is smaller, cheaper, and considerably easier to act on: keep the agent's memory in files you can open, keep its permissions in rules you wrote, and make sure you could move both tomorrow.
Suzent is an open-source, local-first implementation of that idea. The quickstart takes about five minutes, and the source is Apache-2.0.