Memory and interaction in a simulated town
Park and colleagues' 2023 study placed 25 generative agents in a sandbox environment. The architecture combined stored experiences, reflection and planning. The researchers examined the believability of individual behavior and social coordination, including the spread of an invitation to a party. Generative Agents, 2023.
This provides an example of agent architecture and interaction. Believable behavior in that setting does not establish reliable purchase predictions for a business.
Agents grounded in real participants' self-reports
In a separate study, Park and colleagues built agents using interviews, surveys or both from 1,052 Americans. The evaluation included held-out survey responses and used participants' own two-week answer consistency as a benchmark. The current paper distinguishes the results of the different input methods. Self-report grounded agents, revised 2026.
Those conditions matter. A persona generated from a demographic sketch is not equivalent to an agent built from a real participant's interview and survey responses. The reported results are not a general accuracy score for synthetic customers.
Limits of our method
Model-generated reactions can be plausible and still wrong. Several generated personas may repeat the same assumption. Giving a panel evidence that already contains the expected conclusion can also contaminate the result.
Our internal studies have exposed these problems. Source checks and a separate challenge to the findings are part of the method because a persuasive explanation can exceed the available evidence.
The resulting hypotheses still need real-world tests. Founder Effect does not establish market size, conversion rates or willingness to pay from simulated reactions alone.