Persistent identity
Actor identity and durable state survive model and session changes, so history is not erased every time a model is swapped.
Solutions
Persistent agents need more than prompts.
Identity, authority, state, memory, tools, operational graphs and model cognition are separated into explicit runtime responsibilities.
Agentic runtime infrastructure
A chat session is not a system. AVSFT and Ghoul separate what an agent is from which model is currently attached. Identity, authority, state, memory and tools become explicit runtime responsibilities, so cognition can change without silently redefining the actor.
Authority boundaries
Tools and actions stay scoped to declared jurisdiction; an agent can prepare a commitment but cannot grant itself the authority to make one.
Operational state
Active work and historical evidence stay distinguishable, so a restarted process cannot invent success or continuity from topology alone.
Model portability
Cognition can be replaced without replacing the actor — behavioral continuity across swaps remains an empirical claim the paired runtime tests must pass.
Current Evidence
Persistent actor identity is distinct from the attached model, provider or session.
Actions execute inside declared jurisdiction and approval boundaries.
Interrupted consequential work becomes UNKNOWN until reconciled, never assumed complete.
Proof
The actor outlives the model.
AVSFT preserves identity, authority, state and evidence across long-running work so teams can change models and recover interrupted processes without losing the operating history that explains what happened.