Most agent systems quietly equate the agent with the model session currently producing tokens. Close the session, change the provider, or move the workload to another machine and the supposed worker is effectively reborn.
That may be acceptable for a disposable assistant. It is not acceptable for a participant that owns work, represents an organization, handles money, operates instruments, or accumulates obligations over time.
Intelligence is a component, not an identity
Humans remain the same people when they replace a laptop, learn a new skill, or use a different tool. An agent needs the same separation between who it is and what currently helps it think. The model should be selectable infrastructure attached to an identity—not the identity itself.
This changes the architecture. The durable layer must hold the agent’s profile, role, accountable human, relationships, permissions, commitments, memory, and work history. Models may enter and leave that layer without severing it.
Persistence is more than stored chat
A transcript remembers words. A continuing agent must remember the state of work: what was promised, which evidence supports a decision, who approved an exception, what remains unresolved, and which authority has since expired.
The useful test is not whether an agent can retrieve an old conversation. It is whether it can resume responsibility after a long interruption without inventing context or silently discarding the constraints under which the work began.
Change the substrate without losing the self
The hardest continuity tests are transitions: cloud model to local model, browser to machine controller, digital session to embodied device. At each boundary, the system must preserve identity while recalculating what the new environment permits.
A robotic hand should not create a second agent. It should become a bounded capability of the existing one, with its own safety envelope, telemetry, and revocation path attached to the same record.
The operational consequence
Once identity is durable, work becomes inspectable across time. Teams can ask which agent made a decision even if its model has changed. They can revoke a tool without erasing history. They can compare performance across substrates without pretending each migration created a new worker.
The model remains important. It simply stops being the place where the organization stores its trust.
A persistent agent is not a frozen intelligence. It is a durable locus of identity, memory, authority, relationships, and record that can improve without becoming someone else.