GUIDES · ARCHITECTURE
Integrate synthetic populations into an agent architecture.
Agent architectures reason well but ground poorly: when a decision depends on human reactions, the agent imagines populations instead of querying them.
This guide shows how to add a behavioral evidence layer to an agent architecture: synthetic populations with controlled, queryable and traceable properties.
THE DIAGNOSIS
Why do agent architectures fail at decision support?
Agent architectures know how to decompose a problem, orchestrate tools and draft syntheses. Their weakness appears when they must anticipate a human reaction.
Without an evidence layer, the agent queries the LLM that powers it. Left on its own, an LLM tends to produce the most probable reaction and compress the internal variance of groups.
Multiplying agents does not solve the problem: ten agents querying the same model average ten times.
- LLM alone
- Most probable reaction
- Compressed variance
- Minority voices lost
What is missing is not more reasoning capacity. It is a behavioral evidence layer.
THE PRINCIPLE
Separate the reasoning layer from the evidence layer.
REASONING LAYER
The agents
Decompose the decision, formulate hypotheses, orchestrate analyses and draft the outputs.
EVIDENCE LAYER
The synthetic population
Provides the behavioral material: individuals generated under statistical constraints, structured into typologies and queryable individually.
Agents no longer speculate about reactions. They collect them.
- Decision
- Agents
- Synthetic population
- Interviews
- Traceable insights
- Informed decision
The quality of reasoning remains the quality of the agents. The quality of the evidence becomes independently measurable.
STEP 1
Choose the integration pattern.
TOOL
The population is exposed to agents as a callable tool through an API. The agent decides when to query it and receives traceable results.
SKILL
The population and its cognitive models become a capability integrated into the organization's system.
DEDICATED SERVICE
Simulation becomes an isolatable stage in the pipeline, between framing agents and synthesis agents.
The choice depends on three criteria: query frequency, asset ownership and auditability requirements.
STEP 2
Expose the population through typed services.
GENERATION
Instantiate a coherent population using typologies relevant to the decision, calibrated on public data and proprietary segments.
QUERYING
Conduct dynamic interviews, collectively or individual by individual, with contextual follow-ups.
PROJECTION
Calculate behavioral trajectories under multiple scenarios over months or years.
- 7 to 12 follow-ups on average per person.
- Several hundred to more than a thousand compute threads in parallel.
STEPS 3 & 4
Ground the reasoning. Preserve the trace.
- 01 · Frame the decision
- 02 · Build the scenarios
- 03 · Instantiate the population
- 04 · Conduct the interviews
- 05 · Compare the reactions
- 06 · Trace every conclusion
If an insight is surprising, agents can reopen an interview, drill into a typology or test a variant.
No behavioral claim in the final output should exist without a source in the evidence layer.
Each insight retains its source interviews. Each synthesis retains the insights it aggregates. The final output remains challengeable line by line.
- Decision
- Scenario
- Interview
- Insight
- Synthesis
- Recommendation
STEP 5
Govern the whole system.
DATA
No directly identifying data enters the evidence layer. Proprietary data is used for calibration after strict anonymization and does not retrain any model.
SOVEREIGNTY
The evidence layer runs on controlled infrastructure, up to permanent on-premise installation and full isolation for regulated organizations.
RESPONSIBILITY
The architecture informs the decision. It does not make it. The final decision remains human.
STATED LIMITATIONS
Three anti-patterns.
01
The agent simulates its own population.
Every answer may look plausible. The population as a whole remains pulled toward the average.
02
The population has no evidence.
Without measured properties or a calibration report, grounding is only a dressed-up assumption.
03
The decision is delegated.
The architecture sharpens the decision-maker's judgment. It never replaces their responsibility.
FAQ
Frequently asked questions.
How is this different from RAG?
RAG grounds agents in documents: “what does the corpus say?”. A synthetic population grounds them in behavior: “what will people do?”. The two evidence layers are complementary.
What query latency should be expected?
A query wave takes minutes. From the brief to the first delivered dossier, a complete cycle typically takes between 20 and 30 minutes.
Is the architecture tied to a particular LLM provider?
No. The evidence layer is model-agnostic and is exposed by API to any agent architecture.
Where does the evidence layer run?
On controlled French infrastructure for standard deployments, with the option to install it in the client's environment up to full isolation.
Is my organization's data used to train models?
No. It is used exclusively for calibration after strict anonymization. It does not retrain any model and does not benefit any other client.
YOUR NEXT DECISION
Ground your architecture.
The three patterns described in this guide are industrialised in our system.
