An ordinary model call generates output from the current context. An agent loop adds temporal extension: the system observes the situation, constructs state, reasons under goals and policies, acts through an environment, observes the result and updates what happens next.
The proposed structure starts here because every later social-AI problem depends on this loop. Memory only matters because later actions depend on earlier events. Trust only matters because behavior accumulates across time. Autonomy only matters because the agent may choose means, initiate actions or preserve commitments beyond a single prompt.
Agency should be treated as a spectrum. A chatbot responds. An assistant responds and may use tools. A task agent pursues a goal across steps. A persistent agent carries state across sessions. An adaptive social agent also maintains identity, relationship state and governed learning.
