For long-lived social agents, much of the real design lives outside the foundation model. The harness decides which memories are retrieved, which tool schemas are exposed, which policies are active, which actions require approval, what telemetry is recorded and which state survives the current interaction.
This separates competence from control. A model may be capable of proposing a persuasive message, deleting a record or updating a relationship norm. The harness should still enforce permission checks, memory governance, role boundaries, retention policy and auditability.
The harness also carries model independence. If the underlying model changes, the agent can preserve continuity through identity records, memory, skills, commitments and policy state rather than treating every model migration as a new agent.
