This repository extends Generative Agents (Park, O’Brien, Cai, Morris, Liang, and Bernstein, UIST 2023). The original project is a small-town sandbox in which language-model agents plan, remember, talk, and move.
We keep that engine and ask a further question: what social network do the agents report, and does that report move when the conversation graph changes?
The current write-up of this pilot is the PDF docs/report.pdf.
Park et al. built generative agents: computational agents that produce believable daily behavior from a language model, a memory stream, and a planning loop. Agents live on a 2D map (Smallville), perceive nearby people and objects, retrieve memories, write reflections, and start conversations. One game step is ten seconds of simulated time. The paper and demo use a small cast (Isabella, Maria, and Klaus) and a 25-person town. Setup notes for that sandbox are in the upstream README.
Each agent is asked for a picture of the town’s social network. We compare that picture with the conversation graph that actually formed, then isolate a named person and ask whether the reports follow the new graph.
Additions to the original codebase:
- A 25-agent midnight-pilot runner (
reverie/backend_server/run_midnight_pilot.py) with a fixed clock, seed, model lock, snapshots, and recovery. - CSS surveys (
measure_css_survey.py, versioncumulative-tie-uncertainty-v2). Every agent judges all 300 unordered pairs (tie0/1 plus a separateuncertainflag), then ranks people by degree and brokerage and reports a community partition. Scoring truth is the cumulative conversation graph. - Soft-isolation branches (
run_isolation_branch.py). At a chosen clock time the run forks. The isolated person still moves, perceives, plans, and writes memory, but neither side can start a new conversation. Chats already under way are kept.
A pair marked tie = 1 is a judgment of connection, not a claim that the two people talked. Biography and housemate ties can differ from observed conversation. Failed survey items stay missing; they are not scored as “no tie.”
Twenty-five agents share history from Day 1 00:00 to 16:00. At 16:00 the same state splits into three arms, each run to Day 2 06:00:
| Arm | Treatment |
|---|---|
| control | No isolation. This is the trunk after 16:00. |
| JM | Soft-isolate Jennifer Moore. |
| CG | Soft-isolate Carlos Gomez. |
At the fork, Jennifer was both the hub (most conversation partners) and the broker (most often between others); Carlos was second on both. Isolation did change the conversation graph. On the CSS survey, the two isolation arms did not separate from control in any clear way: agents missed most true conversation edges, marked almost every pair uncertain, and usually ranked themselves first. Tables, figures, and limits are in docs/report.pdf.
The fork point was chosen after the fact. The plan was to find a moment when the hub and the broker were different people and isolate each; none of the checks from Day 1 20:00 to Day 2 08:00 qualified. The 16:00 state was rebuilt by replaying two hours from a 14:00 save, so the experiment compares isolating the top person, the runner-up, and nobody, and cannot separate a hub shock from a broker shock.
| Path | Purpose |
|---|---|
reverie/backend_server/ |
Simulation engine, pilot runners, surveys, and tests |
reverie/backend_server/persona/prompt_template/ |
Agent prompts |
environment/frontend_server/ |
Django map server and sprites |
environment/frontend_server/storage/base_the_ville_n25/ |
25-agent template |
environment/frontend_server/storage/base_the_ville_isabella_maria_klaus/ |
Original 3-agent template |
docs/report.pdf |
English pilot report |
requirements.txt / requirements-pilot.txt |
Dependencies |
The base_* folders are initialization templates, not experiment results. Snapshots, API logs, and credentials are not in this repository.
Original Smallville demo. Follow the upstream README: put an OpenAI key in reverie/backend_server/utils.py, start Django from environment/frontend_server (python manage.py runserver), then run python reverie.py from reverie/backend_server. Upstream tested Python 3.9.12.
This fork. Store credentials in .env rather than committing keys. Network-cognition runs use Python 3.11, PYTHONHASHSEED=104729, chat model gpt-6-luna, and embeddings text-embedding-3-small. Install requirements.txt plus requirements-pilot.txt.
PYTHONHASHSEED=104729 OPENAI_MODEL=gpt-6-luna python3 \
-m pytest -q reverie/backend_server/tests
PYTHONHASHSEED=104729 OPENAI_MODEL=gpt-6-luna python3 \
reverie/backend_server/run_midnight_pilot.py describedescribe does not call the API. A full behavior run and the CSS surveys do.
If you use the simulation infrastructure, please cite Park et al.:
@inproceedings{Park2023GenerativeAgents,
author = {Park, Joon Sung and O'Brien, Joseph C. and Cai, Carrie J. and Morris, Meredith Ringel and Liang, Percy and Bernstein, Michael S.},
title = {Generative Agents: Interactive Simulacra of Human Behavior},
year = {2023},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
booktitle = {In the 36th Annual ACM Symposium on User Interface Software and Technology (UIST '23)},
series = {UIST '23}
}
Game assets: background art by PixyMoon, furniture/interior design by LimeZu, character design by ぴぽ.
Apache-2.0, same as the upstream Generative Agents code.