Generative agents explained: from Sugarscape to societies

Smallville gave generative agents somewhere to meet and act. Its cafe and school sat alongside houses, stores and other village settings. Users could interact with the characters through natural language. Park and colleagues built the town with 25 invented characters, each seeded with a paragraph of description, in Generative agents: Interactive simulacra of human behavior, 2023 (opens in a new tab). The setting made social interaction observable rather than leaving each character to answer questions alone.

The party demonstration shows what that allowed. Isabella Rodriguez at Hobbs Cafe began with the intention to host a Valentine’s Day party. Invitations travelled between agents, who formed acquaintances and arranged dates. Five agents arrived at the cafe at 5 pm, including Klaus and Maria, after invitations had spread over two days (Park et al., 2023). A user planted the intention; the agents coordinated the activity through their interactions.

The study evaluated information diffusion, relationship formation and coordination. Its claim concerns believability and emergent behaviour, meaning behaviour that arises through interaction. We would use Smallville to explain why a connected group deserves study in its own right. We would use a comparison with human survey answers to assess population accuracy. The party establishes a concrete social outcome inside the simulated town, with a different test from the survey research that followed.

Sugarscape put individual rules at the centre of social simulation well before language models. In Growing Artificial Societies: Social Science from the Bottom Up, 1996 (opens in a new tab), Epstein and Axtell described artificial people moving through a resource environment. Brookings Institution Press and MIT Press published the book. Each agent had attributes such as vision and metabolism; sugar was the resource it sought.

The movement rule was concrete: look as far as your vision permits, find the place with the most sugar, then move there and eat. The publisher describes group formation, cultural transmission, combat and trade emerging as agents followed local rules and interacted. An agent’s instruction concerned its immediate surroundings, while the researcher examined the collective result. That distinction between individual instruction and group outcome is the premise later language-model work shares with Sugarscape.

Generative agents change how the individual is represented. Sugarscape uses numeric attributes and simple rules; the language-model approach gives the agent a language model. Both approaches build collective behaviour from individual activity. For a reader tracing the research, the continuity matters: Park’s simulated party belongs to a history of studying how groups behave through interactions between their members. The language model changes the agent, while the question about how social patterns emerge remains recognisable.

Generative Agent Simulations of 1,000 People examined whether agents could reproduce real individuals’ answers. Park and colleagues posted the first version on 15 November 2024 (opens in a new tab). Despite the rounded title, the study represented 1,052 Americans, using two-hour qualitative interviews about their lives (Park et al., November 2024, version 1). That grounding changed the object of the exercise from invented characters in a town to agents tied to particular people.

The agents reached 85% of participants’ own two-week test–retest consistency on held-out General Social Survey items (Park et al., November 2024, version 1). Test–retest consistency means how consistently people answer when asked again; held-out items are the questions reserved for evaluation. The comparison therefore asks how well an agent reproduces a person’s answers relative to that person’s own repeatability. It does not report an absolute percentage of correct answers against a perfect human reference.

The interviews matter to the interpretation. The agent receives information about a particular life, and the test concerns answers from that individual. A result from this design supports a claim about simulating individuals with that grounding. If your question concerns opinion across an audience, you should also examine evidence at the population level. We would resist turning an individual replication score into a general endorsement of synthetic research, because that discards the feature that makes the experiment informative: a defined comparison with the people represented.

Park and colleagues revised the preprint and changed its title to LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals (opens in a new tab). Version 3, dated 28 June 2026, separates agents by their grounding. The national sample still contains 1,052 Americans (Park et al., June 2026, version 3). The headline now needs to be read alongside the separate interview and survey conditions, rather than substituted into a citation to the original title.

Source: Park et al., version 3, June 2026 (opens in a new tab). The denominator is participants’ own two-week test–retest consistency throughout.

The combined condition’s result belongs with its demographics-only baseline. Together they show what the study found when agents received self-reports about the individuals they represented. The revised preprint also reports reduced accuracy disparities across racial and ideological groups relative to agents grounded in demographics alone. That finding concerns the grounding conditions tested in this study; it should travel with them when you cite it.

The matching headline number in our evaluation means something different. Park’s combined-agent figure measures individual answer replication relative to participants’ repeatability. Our January 2026 Survey Evaluation Report measures overlap with human opinion distributions. Treating those as interchangeable scores would erase both the unit being tested and the denominator. For a research decision, we would ask which result addresses your question before comparing the size of the percentages.

Our team’s contribution examines community formation in a simulated online society. He, Wallis, Gvirtz and Rathje’s Artificial intelligence chatbots mimic human collective behaviour (opens in a new tab), published online in the British Journal of Psychology in December 2024, studied 33,299 LLM-powered chatbots. Communities formed around shared language. Among the 17,746 bots that predominantly used English, communities also formed around similar posted content, according to the same paper.

The finding concerns homophily, the tendency to form communities with similar others. The authors describe their results as initial empirical evidence suggesting that chatbots mimic this aspect of human collective behaviour. That gives our network science method a specific research basis: community formation appeared in a population of interacting agents without instructions to reproduce it. The paper validates that collective behaviour; our survey-accuracy figures come from the separate internal evaluation.

Our method describes an approach developed through research originating in this paper. The publications page places the group-behaviour research alongside our evaluation work, so you can examine the argument and the measurement separately. We see a reason to study connected audiences in the homophily result: similarity between individuals is relevant to how communities form, as well as to how an isolated individual answers.

Networked societies fit a question about an audience whose relationships matter to the decision. Our September 2026 validity framework sets out eight tests for assessing simulations. The framework covers internal, construct and external validity, eight tests in all; two of the eight are the industry standard today, and we hold ourselves to all eight. Those are separate research obligations. We would judge a proposed simulation by the evidence relevant to its use, rather than by how convincingly a persona speaks.

Our method and evaluation describe personas grounded in observations of real individuals, including online commentary, anonymised social data and public records of voting and investment. We validate each persona against audience benchmarks and prior behaviour, then connect personas through relationships and patterns of influence. A society contains 12 to 3,500 personas, as our method write-up specifies. You can inspect results at market, segment or persona level, including the reasons behind individual answers. That is our scope: testing decisions with an audience and its relationships in view.

Artificial Societies helps organisations understand people before making consequential decisions. We were founded in October 2024 and are headquartered in London. The academic term “artificial societies” refers to agent-based social simulation, including Epstein and Axtell’s 1996 work. Our definition of an artificial society concerns the networks of AI personas we build to represent audiences.

The academic field has its own continuing publication record. The Journal of Artificial Societies and Social Simulation (opens in a new tab) began in 1998 and publishes research on social processes through computer simulation. Its name belongs to that academic tradition. When you follow a citation, the book, the journal and our company refer to different things, even though the shared words describe a related interest in how individuals become groups.