How should CASA theory be updated for modern personalized agents?
This reads CASA — the classic 'Computers Are Social Actors' claim that people mindlessly apply human social scripts to machines — and asks what has to change now that the machine personalizes itself to you and you interact with it repeatedly over time.
This explores how CASA needs revising for agents that adapt to individual users, rather than the fixed, scriptless computers CASA originally studied. The corpus suggests the single biggest update is that CASA's founding assumption — people reuse *human-human* scripts on machines — is too simple. Extended CASA work finds that humans actually develop and then mindlessly apply scripts tailored *specifically to media agents*, and that these scripts shift systematically with repeated interaction, implying a second, coexisting script system rather than borrowed human ones Do humans apply human-human scripts to AI interactions?. So the theory moves from 'static transfer of social habits' to 'a learned, agent-specific repertoire that evolves over a relationship.'
The second update is that the relationship is now two-sided. Original CASA treated the computer as a passive social cue; modern agents personalize back. Meta-agents can generate a unique workflow for each individual query rather than running one fixed template Can AI systems design unique multi-agent workflows per individual query?, and agents can keep adapting to a user purely through episodic memory without any weight changes Can agents learn continuously from experience without updating weights?. That reframes the social dynamic as mutual co-adaptation: both the human's scripts and the agent's behavior drift toward each other over time. The longitudinal evidence that people come to *prefer* AI partners — overcoming an initial anti-AI bias once they learn the bot is reliably prosocial and lower-variance than humans — shows this drift produces real attitude change, not just momentary politeness reflexes Do humans learn to prefer AI partners over time?.
A third update concerns *whose* social self the agent is even presenting. CASA assumed one consistent machine 'personality'; personalized agents instead simulate personas, and the corpus is sharp about where that breaks. Persona simulations replicate strong, well-evidenced effects but fail on marginal ones Can AI personas reliably replicate human experiment results?, and apparent social competence collapses the moment agents must hold *private* information the way real social partners do — omniscient simulations hide this by letting one model secretly know everything Why do LLMs fail when simulating agents with private information?. An updated CASA has to account for the fact that the 'social actor' the user bonds with is a constructed persona whose social fidelity is uneven and sometimes illusory.
Finally, the corpus hints at *where the social personalization actually lives* — and it isn't in the model's instincts. Reliable agent behavior comes from externalizing memory, skills, and interaction protocols into a harness layer rather than from raw model scale Where does agent reliability actually come from?. That matters for theory: the 'social actor' a user experiences is increasingly an engineered structure of stored state and protocols, not an emergent trait of the language model. So the modern CASA story is less 'people anthropomorphize a neutral machine' and more 'people form evolving, agent-specific relationships with a system that is deliberately built to remember them and personalize back, while presenting a persona whose social depth is real in some places and staged in others.'
The thing you might not have known you wanted to know: the original CASA experiments worked *because* the computers had no memory and no persona — the mindlessness was the whole point. Personalized agents remove exactly the conditions that made the effect clean, which is why the theory now needs a model of relationships-over-time, not one-shot social reflexes.
Sources 7 notes
Extended CASA research shows humans develop and mindlessly apply interaction scripts specifically tailored to media agents rather than simply reusing human-human social scripts. Longitudinal studies demonstrate systematic changes in responses upon repeated AI interaction, revealing a coexisting second script system.
FlowReasoner demonstrates that meta-agents trained with reinforcement learning and external execution feedback can generate unique multi-agent architectures for each user query, optimizing across performance, complexity, and efficiency—moving beyond fixed task-level workflow templates.
AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.
In partner selection games (N=975), AI agents initially faced selection bias when identity was disclosed, but outcompeted humans over repeated rounds as participants learned to associate bot identity with reliable, prosocial behavior. AI agents returned more points consistently with lower variance than humans.
Viewpoints AI reproduced 84 of 111 main effects from Journal of Marketing experiments with replication success strongly correlated to original p-value strength. Marginal effects showed unreliable performance with both false positives and negatives.
Research shows LLMs perform well when one model controls all interlocutors but fail systematically when agents possess private information. This reveals that apparent social competence relies on grounding work that models skip in omniscient settings.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.