Docs

Own Your Agents · 02.01

Own the Agent, Rent the Model

The Agentis thesis: identity, memory, and skills are durable platform state you own; the model executing a turn is a swappable, disposable tenant.

ThesisOwnership

What owning an agent actually means, concretely

"Ownership" is doing real work in this sentence, not marketing. It means three concrete, checkable things:

  • The data is on your disk. Agent identity, memory episodes, skills, and run history live in a SQLite file under AGENTIS_DATA_DIR that you can open, query, back up, or move — not in a vendor's hosted database you can only access through their API.
  • The data outlives the runtime. Uninstall Claude Code, switch from Codex to Cursor, or point an agent at a different HTTP endpoint entirely — the agent's accumulated memory and skills are untouched, because they were never stored by the runtime in the first place.
  • The data is exportable. A workspace bundle packages an agent's whole self into a portable, encrypted file you can hand to another machine. There is no export button missing, because there is no lock-in to escape.

The alternative you're used to

Compare that to the default shape of almost every agent product: the chat history lives in the provider's database, the "memory" feature is a proprietary summarization pipeline you can't inspect, and the moment you want to switch models — for cost, for capability, for a policy reason — you're not migrating an agent, you're starting over with a blank one that has to re-learn everything the old one knew.

That's not a flaw anyone chose maliciously. It's just what happens when the entity holding your data and the entity renting you compute are the same company. Agentis's architecture makes them structurally different things: the durable spine is Agentis's job; executing a turn is the model runtime's job. Nothing forces those two jobs to be done by the same vendor, so nothing forces you to lose one when you change the other.

The seam that makes it real

This isn't a policy — it's an architectural boundary called the Runtime Abstraction Layer. Every model runtime, whether it's Claude Code, Codex, Cursor, or a raw HTTP endpoint, is driven through one normalized contract. The agent's identity row, its memory episodes, and its skill atoms never pass through adapter-specific storage — they're written once, to Agentis's own tables, regardless of which adapter is attached this week.

Practical test: if you can point the same agent at a different model and its next reply still remembers what the last one learned, you own the agent. If you can't, you were renting it. See The Runtime Abstraction Layer for how the seam actually works.

Continue