What are simulated agents in surveys? Definition, how they work, and how they differ from synthetic respondents
Simulated agents are autonomous AI entities that mimic human decision-making, memory, and social behavior in research environments. Learn what they are, how they differ from synthetic respondents, the Stanford generative agents research, and where they're useful.
Simulated agents are autonomous AI entities that interact dynamically with each other and their environment, maintaining memory, making plans, and evolving over time. Unlike synthetic respondents (which generate one-shot survey answers), simulated agents have persistent state and emergent behavior that allows them to model complex group dynamics, market behavior, and social interactions. This guide defines simulated agents, explains how they differ from synthetic respondents, summarizes the Stanford generative agents research that established the field, and covers where simulated agents are useful in market research today.
Frequently asked questions
What are simulated agents in market research?
Simulated agents are AI-powered virtual entities that autonomously act, interact, and evolve in a simulated environment to model how real humans might behave in similar situations. In market research and surveys, they are used to predict consumer responses to new products, simulate the spread of information or behaviors through social networks, model markets at scale, and test interventions in environments where real-world experiments would be impractical or unethical. The key distinction from synthetic respondents is that simulated agents have persistent memory, planning capabilities, and the ability to interact with other agents, producing emergent behaviors that no single agent was explicitly programmed to exhibit.
How are simulated agents different from synthetic respondents?
Simulated agents and synthetic respondents both use AI to model human behavior, but they work very differently. Synthetic respondents are stateless: each query produces an isolated response with no memory of previous interactions. Simulated agents are stateful: they have persistent memory, can plan future actions, and interact with other agents over time. Synthetic respondents are best for survey-style data at scale. Simulated agents are best for modeling dynamic systems, group dynamics, and emergent behavior. They are complementary tools, not competitors.
How do simulated agents work?
Simulated agents work through four core components. First, a large language model (LLM) generates the agent’s reasoning and responses. Second, a memory stream stores the agent’s experiences (what happened, who they interacted with, what they observed). Third, a reflection mechanism synthesizes memories into higher-level insights about the agent’s situation, preferences, and plans. Fourth, a planning system uses the agent’s reflections to decide what to do next. When agents interact, their actions become memories for other agents, creating feedback loops that produce emergent behavior.
What is the Stanford generative agents research?
The Stanford generative agents research is the foundational academic work that established simulated agents as a viable approach for modeling human behavior. The 2023 paper “Generative Agents: Interactive Simulacra of Human Behavior” (by Park et al., Stanford and Google) created a virtual town called Smallville with 25 AI agents who developed emergent social behaviors like planning a Valentine’s Day party from a single seed prompt. A 2025 follow-up study scaled the approach to 1,000 agents, finding that simulated agents matched real human survey responses at approximately 85% accuracy on personality, behavioral, and experimental benchmarks, outperforming earlier approaches and showing reduced bias compared to demographic-only models.
Are simulated agents accurate enough for real research?
Simulated agents can match real human responses at 85% accuracy on certain benchmarks, according to the Stanford 2025 1,000-agent study. This accuracy applies to personality assessments, behavioral predictions, and structured experimental outcomes. Accuracy is lower for novel scenarios outside the training data, emotionally complex situations, and individual-level prediction. Simulated agents are most useful for hypothesis generation, system-level modeling, and exploring “what if” scenarios, with real participant validation required for high-stakes decisions.
What are simulated agents used for in surveys and research?
Simulated agents are used for five main purposes in research. First, predicting how groups respond to new products or interventions before real testing. Second, modeling how information, opinions, or behaviors spread through social networks. Third, testing survey designs by running drafts through agent populations to identify confusing questions or response biases. Fourth, simulating hard-to-reach or sensitive populations where direct research is impractical. Fifth, exploring “what if” scenarios at scale, like how different pricing strategies might affect consumer behavior across diverse market segments.
How simulated agents work
Simulated agents are more sophisticated than synthetic respondents because they have internal architecture that mimics aspects of human cognition. Understanding the architecture helps you evaluate when simulated agents are appropriate.
The four core components
1. Language model core. A large language model (typically GPT-4, Claude, Gemini, or similar) provides the agent’s reasoning and language abilities. The LLM is what generates the agent’s “thoughts” and “responses” to situations.
2. Memory stream. Every experience the agent has, every conversation, every observation, gets recorded in a memory stream. Memories are timestamped and tagged with importance scores. When the agent needs to make a decision, it retrieves relevant memories from this stream.
3. Reflection mechanism. Periodically, the agent processes its accumulated memories to generate higher-level reflections: insights about its preferences, patterns it has noticed, conclusions it has drawn. Reflections become memories themselves, which means the agent’s understanding of itself evolves over time.
4. Planning system. The agent uses its memories and reflections to plan what to do next. Plans can be high-level (“I want to organize a party”) or detailed (“I’ll talk to Alice at 3 PM and invite her”).
How agents interact
When two simulated agents interact, the interaction is recorded as a memory for both. Each agent updates its understanding based on what the other said or did. Over time, social structures emerge: friendships form, conflicts arise, information spreads, group decisions get made. None of this is scripted; it emerges from the agents pursuing their individual goals and interacting with each other.
This is fundamentally different from how synthetic respondents work. A synthetic respondent answering a survey question has no awareness of any other respondent. A simulated agent in a population is shaped by every interaction it has had.
Simulated agents vs synthetic respondents
The most important distinction in this space is between simulated agents and synthetic respondents. They are often conflated but solve different problems.
| Dimension | Simulated agents | Synthetic respondents |
|---|---|---|
| Core function | Multi-agent systems that interact, remember, and plan | Single AI personas generating isolated responses per query |
| State management | Persistent memory; reflection; emergent behaviors | Stateless; fresh start each query |
| Time dimension | Temporal; agents evolve over time | Snapshot; no temporal evolution |
| Interaction | Agents interact with each other, creating feedback loops | No interaction; each respondent is isolated |
| Best for | System-level modeling, group dynamics, emergent behavior | Survey responses, scaled feedback, persona-based input |
| Output type | Behavioral traces, dynamic patterns, emergent outcomes | Survey-style aggregated data |
| Computational cost | High (multiple LLM calls per agent per cycle) | Low to moderate |
| Complexity | High (architecture, orchestration, debugging) | Lower (prompt engineering primarily) |
| Maturity | Early; foundational research from 2023-2025 | More mature; vendor platforms widely available |
| Real-world accuracy benchmark | 85% match on Stanford 1,000-agent study | 85-95% on quantitative behavioral patterns |
| Risk profile | Sensitive to prompt design, model bias, simulation drift | Sensitive to sycophancy, training data limits |
| Example use case | Modeling how a product launch propagates through a market | Pre-testing a survey before sending to real participants |
When to use which
Use synthetic respondents when:
- You need survey-style data quickly
- Your research question is structured (rate, rank, choose)
- You are pre-testing a survey before fielding to real respondents
- You need scale for hypothesis generation
Use simulated agents when:
- You need to model how behavior spreads through a network
- You are studying group dynamics or social interaction
- You need to simulate a system over time
- The research question involves emergent behavior
Use both in combination when:
- Synthetic respondents generate initial hypotheses about audience preferences
- Simulated agents test how those preferences play out in interactive contexts
- Real participant research validates the most important findings
Most mature research programs use synthetic respondents for early survey work and simulated agents for systems-level questions, then validate critical findings with real participants.
The Stanford generative agents research
The foundational academic work in this field comes from Stanford University. Two key studies define what simulated agents are and what they can do.
Smallville (2023)
The 2023 paper “Generative Agents: Interactive Simulacra of Human Behavior” by Park et al. created a virtual town called Smallville with 25 AI agents. Each agent had a name, occupation, personality, and relationships with other agents. They lived in a virtual environment with houses, shops, and public spaces.
The researchers gave the agents a single seed prompt: one agent wanted to plan a Valentine’s Day party. From this single prompt, the agents organically:
- Spread word about the party through conversation
- Coordinated to invite each other
- Showed up at the right time and place
- Developed new relationships during the event
None of this was scripted. The party emerged from the agents pursuing their individual goals while interacting with each other. The paper demonstrated that LLM-powered agents with memory, reflection, and planning could produce believable social behavior at scale.
The 1,000-agent study (2025)
The 2025 follow-up study extended the approach to 1,000 simulated agents and tested whether their responses matched real human survey data. Each agent was constructed from a real person’s interview transcript, creating a synthetic version of that individual.
The findings:
- Simulated agents matched real human responses at approximately 85% accuracy on personality, behavioral, and experimental benchmarks
- The approach outperformed earlier demographic-only modeling
- Bias was reduced compared to predictions based on demographics alone
- Simulated agents showed cognitive realism beyond what scripted bots could achieve
This study moved simulated agents from interesting research demo (Smallville) to a serious approach for behavioral prediction at scale.
Why this matters for market research
The Stanford research validates two important claims. First, simulated agents can produce believable individual behavior that matches what real people would do in similar situations. Second, when you scale to populations of agents, emergent dynamics appear that mirror real-world social behavior. For market researchers, this opens up new possibilities for modeling consumer behavior, product diffusion, and social influence at a level of detail that was previously impossible without expensive longitudinal field studies.
Use cases for simulated agents in research
Simulated agents add unique value in scenarios where dynamic interaction matters more than survey-style responses.
1. Modeling market and product diffusion
How does a new product spread through a market? Traditional research can answer this in retrospect (after launch) but struggles to predict it in advance. Simulated agents can model populations of consumers with different attitudes, social networks, and decision rules, then run simulations to see how a product would spread. The output is not just a forecast but a behavioral trace showing which agents adopted first, who influenced whom, and where adoption stalled.
2. Simulating social influence and opinion dynamics
Research on opinion change, persuasion, and social influence has historically required field experiments that are expensive and ethically constrained. Simulated agents can model how opinions form and spread in a population, how minorities can influence majorities, and how information cascades unfold. These models can inform marketing strategy, public health communication, and policy design.
3. Pre-testing interventions before real-world deployment
Before launching a new pricing model, marketing campaign, or product change, simulated agents can model how a population of consumers would respond. This pre-testing identifies obvious failures, surfaces unintended consequences, and helps refine the intervention before real-world deployment. It is most useful for interventions that affect groups dynamically rather than individuals in isolation.
4. Modeling hard-to-reach or sensitive populations
Some research populations are difficult to recruit ethically or practically: vulnerable patients, executive decision makers, illegal market participants, populations in crisis. Simulated agents can model these populations for early-stage exploration, with the understanding that real validation is required before high-stakes decisions.
5. Stress-testing surveys and research designs
Simulated agents can run through draft surveys to identify confusing questions, response biases, and skip logic errors. They are similar to synthetic respondents in this use case, but their stateful nature means they can test surveys that have multiple sessions or longitudinal designs.
Limitations of simulated agents
Simulated agents inherit the limitations of LLMs and add new ones from their architectural complexity.
1. Sensitivity to initial conditions
Small changes in agent setup, prompts, or environment can produce large differences in outcomes. This is similar to chaos in real complex systems but makes simulations harder to reproduce and trust. Researchers using simulated agents must run many simulations and report distributions of outcomes, not single runs.
2. Hallucination and drift
LLMs can hallucinate facts, contradict themselves, or drift away from their persona over time. In long simulations, agents may develop inconsistencies that undermine the validity of the results. Mitigation requires careful prompt engineering and frequent state validation.
3. Training data limitations
Like synthetic respondents, simulated agents are shaped by the LLM’s training data. Behaviors and patterns underrepresented in training data are poorly modeled. This is especially limiting for niche populations, novel behaviors, and cultural contexts outside the training distribution.
4. Computational cost
Running thousands of simulated agents over many cycles requires substantial LLM compute. Each agent may require multiple LLM calls per simulation step (for memory retrieval, reflection, planning, interaction). Costs scale rapidly with agent count and simulation length.
5. Validation challenges
How do you know if a simulation is accurate? Validating simulated agents against real-world outcomes is difficult, especially for novel scenarios where ground truth does not yet exist. The Stanford 1,000-agent study used real survey data as ground truth, but most real-world simulations cannot validate this directly.
6. Ethical and interpretability challenges
When simulated agents produce unexpected or troubling behaviors, it can be unclear whether these reflect real human dynamics or artifacts of the simulation. Researchers must be careful not to over-claim insights from simulations and must subject findings to rigorous validation before acting.
Tools and platforms
The simulated agents space is early but evolving fast. The current landscape includes:
| Category | Examples | Best for |
|---|---|---|
| Academic/research frameworks | Stanford’s generative agents codebase, AgentVerse, CAMEL | Research and experimentation |
| Behavior modeling platforms | Aaru, Decagon, generative agent startups | Commercial behavioral simulation |
| Multi-agent orchestration | LangGraph, CrewAI, AutoGen | Building custom agent simulations |
| Agent testing platforms | Sierra, Maxim AI, LangWatch | Testing AI agent products (different use case but related) |
Most commercial simulated agent platforms are early-stage. Vendor evaluation should include questions about:
- What LLM is the agent powered by, and can you choose?
- How is memory implemented, and can you inspect it?
- What validation has the platform done against real-world outcomes?
- What is the cost model for running large simulations?
- How does the platform handle reproducibility and simulation runs?
How to use simulated agents responsibly
The early stage of this field demands careful practice to avoid over-claiming or making poor decisions based on simulation outputs.
1. Treat simulations as hypotheses, not conclusions
A simulated agent run produces hypotheses about what might happen. It does not produce conclusions. Always validate critical findings with real research before acting.
2. Run many simulations and report distributions
A single simulation run is one possible outcome from a stochastic process. Run 100+ simulations and report the distribution, not just one trajectory. Highlight outliers and surprising behaviors for investigation.
3. Document model versions, prompts, and seed conditions
Simulations are sensitive to initial conditions and model versions. Document everything so you can reproduce results and explain differences when models change.
4. Validate against real data wherever possible
If you have real survey data, behavioral data, or experimental data on the question you are simulating, calibrate your simulation against it. The Stanford 1,000-agent study is a good model: build agents from real data and validate against real responses.
5. Communicate uncertainty to stakeholders
When presenting simulation results, communicate the limitations clearly. Stakeholders should understand they are looking at model output, not measured human behavior, and should treat findings accordingly.
6. Combine with other methods
Simulated agents are most powerful when combined with synthetic respondents, real participant research, behavioral data, and expert input. No single method is sufficient for important decisions.
The future of simulated agents in research
Simulated agents are at the same stage that synthetic respondents were 18 months ago: rapidly advancing, intensely debated, and beginning to find real use cases. Three trends will shape the next 18 months:
1. Validation studies will multiply. More academic and industry studies will compare simulated agent outputs to real-world outcomes, building empirical confidence in where simulations work and where they fail.
2. Integration with traditional research will become standard. Mature research programs will use simulated agents for early hypothesis generation and systems-level questions, with real participant research for validation, similar to how the field is using synthetic respondents today.
3. Domain-specific platforms will emerge. General-purpose simulated agent frameworks will give way to platforms tailored to specific use cases: consumer behavior modeling, social influence simulation, market dynamics, organizational behavior.
For teams evaluating this space, the synthetic respondents guide covers the closely related but distinct technology of survey-style AI respondents, and the AI in user research guide provides broader context on how AI tools are reshaping research practice. Simulated agents are not a replacement for understanding real humans; they are a new tool for exploring questions that would otherwise be impossible to answer.