In a social agent, reasoning is not just answer generation. The agent must interpret ambiguous situations, select relevant memories, infer possible goals, plan under constraints, notice uncertainty and decide when to ask rather than act.
World models represent how the agent believes the environment works. Social world models add assumptions about people, roles, norms, timing, commitments and consequences. These assumptions should remain corrigible: the agent can use them to reason, but should not treat weak social inference as settled knowledge.
Reflection converts episodes into higher-level patterns. It can support better planning and identity continuity, but it can also create false theories about the user or the relationship. Metacognition is therefore part of safety: the agent should know when it is hypothesizing, when evidence is thin and when user validation is needed.
