This post is the third in a series exploring reflection techniques in AI systems. For the complete series, see our posts on Reflexion and Self-Refine. The rest of the series will be published shortly.

Can virtual characters wake up, make breakfast, chat with neighbors, and organize a party without a human pulling the strings? Recent research from Stanford and Google introduced generative agents – computational software agents that simulate believable human behavior1. These AI-driven characters inhabit a virtual town, carrying out everyday activities and social interactions on their own. The concept merges large language models (like the tech behind ChatGPT) with simulated game agents, resulting in AI characters that act uncannily human-like in their decisions and dialogues.

What Are Generative Agents?

Generative agents are essentially AI personas with memories, motivations, and the ability to plan. Unlike traditional video game NPCs (non-player characters) that follow pre-programmed scripts, generative agents rely on a generative model (a large language model) to decide what they do and say based on their experiences2. Each agent is given an initial background (a short biography of identity, job, relationships, etc.), and from there the agent runs autonomously. It observes its environment, remembers interactions, and generates new behaviors appropriate to the situation. In other words, the agent “writes” its own script as it lives its virtual life, resulting in more dynamic and believable behavior than hard-coded AI characters.

How the Stanford Experiment Worked

To demonstrate generative agents, researchers populated a sandbox world (inspired by The Sims) with 25 unique agents1. The environment – nicknamed “Smallville” – contained houses, a park, a cafe and other familiar settings for the agents to navigate. Each agent started with a one-paragraph description (e.g. name, occupation, and a few traits) and then improvised the rest using the AI model2. Throughout a two-day simulation, the agents exhibited strikingly human-like routines and interactions: they woke up in the morning, went to work, socialized with others, formed new friendships, and even gossiped about events in town.

Behind the scenes, generative agents have a special architecture to support their believability. Key components include:

  • Memory stream: As an agent experiences events, it continuously logs them in natural language (for example, “John heard about the Valentine’s Day party from Isabella.”). This running memory lets the agent recall past details later1.
  • Reflection: Agents periodically pause to synthesize their recent memories into higher-level insights. For instance, an agent might reflect “I’ve been chatting a lot with Sam lately; we’re becoming friends.” These reflections help the AI form longer-term attitudes and plans1.
  • Planning: Based on its current needs, goals, and relevant memories, an agent devises plans for what to do next. It might decide to go eat lunch, start a conversation, or prepare for an upcoming event. This planning step uses both the agent’s scripted knowledge (its initial profile) and its accumulated experiences to choose realistic actions1.

Using this setup, the Stanford team observed emergent social behavior. For example, the researchers planted one idea: one agent wanted to host a Valentine’s Day party. From that single seed, the agents autonomously spread invitations to friends, who told friends of their own; some agents asked each other out on dates to the party, and ultimately many showed up at the right time and place for the event12. All of this happened without explicit instructions for any of those intermediate steps – the coordination emerged from the agents’ interactions. In another scenario, one agent decided to run for town mayor, spurring political discussions among the others about the campaign2. Impressively, when human evaluators later compared these AI agents’ responses to those of real people pretending to be the agents, the AI-driven agents were rated as more believable and human-like than the humans2. This highlights how consistent and contextually rich the generative agents’ behavior was. (The research team has even released an interactive demo replay and open-sourced the code for Smallville, so others can explore these agents in action3.)

Why Do Generative Agents Matter?

Generative agents represent a significant step forward in AI for interactive worlds and simulations. First, they promise more lifelike NPCs in games. Game developers are intrigued by the possibility of characters that truly think and react on their own, making gameplay more immersive2. Instead of repetitive canned dialogues, future games could feature villagers, allies, or even opponents with distinct personalities that remember the player’s actions and dynamically change their behavior.

Beyond entertainment, generative agents open doors for social simulations. Scientists can use these believable agents as virtual people to prototype and study complex social dynamics. For instance, researchers could simulate how a community of AI agents responds to a new policy or an emergency, offering insights into how real groups might behave in similar situations2. This kind of “social prototyping” lets us safely explore scenarios (e.g. responses to misinformation or pandemic measures) in a controlled virtual setting before trying them in the real world.

Finally, the work on generative agents is important for the broader goal of building AI that operates in open-ended environments. Today’s AI assistants and chatbots typically reset after each interaction or follow narrow scripts. In contrast, generative agents show how we can create AI systems that persist over time, accumulate knowledge, interact with each other, and adapt to unexpected situations – all characteristics we’d expect from human-like intelligence. By integrating long-term memory and reflection into an AI’s architecture, the Stanford study demonstrates a path to more autonomous and human-centric AI. These agents hold up a mirror to human behavior, but also highlight new challenges (like ensuring AI “societies” remain aligned with human values). Overall, generative agents are a compelling development in understanding how artificial minds might one day live and learn alongside us in richly simulated worlds.

  1. Park, J. S., et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. arXiv:2304.03442 [cs.AI]. ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  2. Stanford HAI. (2023). Computational Agents Exhibit Believable Humanlike Behavior. Stanford HAI. ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  3. Park, J. S., et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. Project Website. ↩