Integrating synthetic populations into an agent architecture.

Agent architectures excel at decomposing a problem, orchestrating tools and writing syntheses. They fail on one precise point: behavioral grounding. When an agent must assess how customers, citizens or teams will react to a decision, it has no resource other than asking the LLM that powers it. Yet an LLM left to itself has a structural drift: it averages. It produces the most probable reaction and flattens the internal variance of the groups; rare combinations and minority voices disappear, precisely what the decision needs to see. Multiplying agents corrects nothing: ten agents querying the same model average ten times.

The problem is not the reasoning capacity of the architecture. It is the absence of an evidence layer: a substrate the agents can interrogate, and which itself possesses controlled statistical properties and measured deviations. That is exactly the role of a synthetic population.

The architectural principle is a separation of layers. The reasoning layer, the agents, decomposes the decision, formulates the hypotheses, orchestrates the analyses and writes the outputs. The evidence layer, the synthetic population, provides the behavioral matter: modeled individuals up to several million, generated under multidimensional statistical constraints, structured into granular typologies, interviewable one by one. The agents no longer speculate about reactions: they collect them, through interviews conducted on the population, and return them with their trace.

This separation changes the nature of the guarantees. The quality of the reasoning remains that of the agents; the quality of the evidence becomes measurable independently: deviations from the reference distributions documented in a calibration report, variance compared against survey references, traceability of every insight down to the individual interviews. A grounded agent architecture is auditable where a bare agent architecture is not.

Three patterns cover most cases. The tool pattern, first: the population is exposed to the agents as a callable tool, through an API, just like a document base or a computation engine; the agent decides when to interrogate the population, formulates its requests, receives traced results. It is the fastest pattern to deploy and the most common. The skill pattern, next: the population and its cognitive models are integrated as a capability of the LLM itself, held under exclusive license or full ownership by the organization; it is the pattern of players who turn simulation into an internal strategic asset. The dedicated service pattern, finally: the simulation becomes a full step of the decision pipeline, orchestrated between the framing agents upstream and the restitution agents downstream; it is the pattern of governance architectures where each step must be isolable and auditable.

The choice rests on three criteria: interrogation frequency (the tool suffices for occasional uses, the skill and the dedicated service serve continuous uses), the ownership requirement (the skill answers proprietary-asset strategies), and the auditability requirement (the dedicated service isolates the simulation for control functions).

Whatever the integration, the population is exposed to the agents through three families of services. Generation: instantiating a coherent population according to the typologies relevant to the decision, calibrated on public data and on the organization's proprietary segments. Interrogation: conducting dynamic interviews on the population, collectively or individual by individual, with contextualized follow-ups, 7 to 12 per person on average, run in parallel across several hundred to more than a thousand compute threads. Projection: computing the behavioral trajectories over months or years under the scenarios formulated by the agents, with the complete trace of the reasoning.

Interface discipline matters as much as the services: every response returned to the agents carries its traceability metadata, the identifiers of the source interviews, the typologies mobilized, the calibration deviations of the population interrogated. This is what will make it possible, at the end of the chain, to trace an agent's recommendation back to the interviews that ground it.

The typical workflow runs as a loop. The framing agents turn the decision into a simulatable question and contrasted scenarios. The population is instantiated at the relevant scale. The agents submit each scenario to the population through interviews; the insights come back with their traces. The synthesis agents confront the scenarios, identify the tipping points and write the output, each conclusion citing its source interviews. If an insight surprises, the agents can follow up on the interview, drill into a typology, test a variant: the loop is iterative, and each iteration counts in minutes of compute, not in weeks of fieldwork.

The golden rule of grounding: no behavioral claim in the final output that does not come from the evidence layer. What belongs to the agents' reasoning, decomposition, hypotheses, presentation choices, is owned as such; what belongs to human reactions is systematically sourced in the interviews. This editorial discipline is what separates grounded decision support from eloquent writing.

The risk specific to agent chains is provenance laundering: through summaries of summaries, the final output no longer knows where its claims come from. The remedy is architectural: traceability propagates as a mandatory metadata at every link. Every insight keeps its interview identifiers; every synthesis keeps the identifiers of the insights it aggregates; the final output is contestable line by line. Risk and compliance departments can audit the complete chain, from the delivered file down to the individual interviews, and check the calibration deviations of the population mobilized. This is the condition for an agent architecture to be admissible in regulated decision processes.

Three governance requirements structure the deployment. Data first: no personal data enters the evidence layer, the populations are rebuilt from anonymous aggregates and segments, natively compliant with the RGPD; the organization's proprietary data serves calibration after strict anonymization and retrains no model. Sovereignty next: the evidence layer is hosted on controlled infrastructure, in France for standard deployments, on-premise, up to complete air-gapped isolation, for regulated players; the agent architecture can thus satisfy requirements that dependencies on foreign hyperscalers rule out. Responsibility finally: the architecture illuminates the decision, it does not make it; the positioning with regard to the AI Act is that of a traceable decision-support instrument, where the final decision remains human.

The first anti-pattern is the agent that simulates its own population: asking the LLM to play the respondents, with no statistical constraints and no measured deviations. Every answer looks plausible; the population as a whole is false, pulled toward the average. The second is the population without proofs: an evidence layer whose properties are neither measured nor documented adds only the appearance of rigor; without a calibration report, the grounding is a dressed-up hypothesis. The third is delegation: wiring the agent chain directly into the execution of the decision. An organization that delegated its decisions to an agent architecture, however well grounded, would commit a major methodological and political fault. The architecture augments the decision-maker's lucidity; it does not replace their responsibility.