Identity is more than memory.
It includes permissions, tools, policies, history, goals, and a record of what the agent has done.
Claude or Codex can be the brain you turn on and off.
The agent is the persistent layer around it.
Most agents today donβt really exist outside the session.
Close the window and the context is gone. New session, new identity.
That makes the model useful, but the agent disposable.
The model can be replaced.
The identity should persist.
A developer agent produces a change. A QA agent checks it against the reference and returns feedback. The developer tries again.
The reference stays stable. The path to it doesnβt.
Thatβs why I see agentic engineering as a stochastic control problem.
AI agents are stochastic. Stop trying to make them deterministic.
Make the target deterministic: requirements, acceptance criteria, tests, SonarQube quality gates.
Then close the loop:
build β evaluate β feedback β retry
You donβt control every action. You control the loop.