A Population You Can Survey
Twelve general-purpose agents for a public-opinion research company’s digital-persona survey system, from building the personas to reporting what they say.
The problem.
Public-opinion research runs on field panels. Recruit people, field a questionnaire, wait, weight the sample, publish. Every additional question costs money and weeks, which is why a survey asks few questions, and asks them late.
The shortcut AI appeared to offer, asking a language model what people think, falls apart on inspection. One model has one voice. It cannot represent how a retired teacher differs from a rideshare driver, and it will produce a confident national average that belongs to nobody.
Our client is building the answer: a societal simulation system built on digital personas, which can be surveyed the way a real panel is. It is built in modules, and the agents for several of them were ours to build.
The system.
We built twelve general-purpose agents for the system, in three groups. Each does one job, through a defined interface, inside the client’s platform.
Personas and surveys. The Digital Persona agent builds representative personas from social media data. The Representation Criteria agent balances their demographic, political, and psychological spread. The Surveyor agent puts questions to them, and the Response Simulation agent answers as each persona most probably would.
Analysis. The Sentiment agent reads the emotional range of a person’s posts, and the Topic Modeling agent finds the themes they return to. The Consistency Test agent checks every answer for logical and semantic coherence. The Dashboard agent turns the results into a visual report.
Population insight. The Opinion Flow agent tracks how a persona’s views shift over time. The Collective Insight agent draws shared trends and dividing lines from many answers at once. The Distribution and Variance agent checks answers across demographic and psychometric breakdowns, and the Style Refinement agent keeps each persona’s own voice in what it writes.
In operation.
Each agent went through the same five steps: design, prototype, MVP, production, and monitoring, with a working deliverable at every step. The groups were delivered one after another. Each was tested end to end with the agents already in the system, so the twelve work together as well as alone.
The first two groups, eight agents, run in production. The third group has reached stable releases and is in final integration.
Ready to build your AI capability?
We accept a limited number of research partnerships per quarter to ensure high-fidelity delivery.
Book a Discovery CallNot ready for a call? contact@whitehand.ai