ACM SIGSPATIAL 2026 · Demo

Urban simulation you can inspect, resume, and validate.

En-AgentSociety is an open-source framework for reproducible, regional-scale simulations with LLM-driven generative agents.

Accepted as a demo at the 34th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems.

The demo

From plausible narratives to research evidence.

En-AgentSociety records the information needed to debug agent decisions, compare generated mobility with empirical data, and reproduce long-running regional simulations. It extends the original AgentSociety framework with observability, fault tolerance, and validation workflows.

The observability stack maps a simulated daily mobility routine into high-level modules and a timestamped 10-minute workflow. In the web UI, each workflow icon opens the prompt-response pair for that timestamp, making the agent's LLM interactions directly debuggable.

Diagram showing agent state, a daily event timeline, and behavioral blocks

What it adds

Built for credible generative-agent experiments.

Full observability

Inspect prompts, responses, memories, agent states, trajectories, latency, and token usage throughout a run.

Mobility validation

Export trajectories and semantic visits for comparison with empirical datasets, mobility laws, motifs, OD matrices, and activity distributions.

Resumable experiments

Persist agent memory, messages, mobility state, and simulator state so interrupted experiments restart safely.

Regional workflows

Combine CitySim-style behavioral modules, traffic simulation, scalable maps, and flexible DuckDB or ClickHouse storage.

Get started

Run your first experiment.

En-AgentSociety requires Python 3.11 or newer. Start with the example configuration, then inspect the results in the web UI.

$ git clone https://github.com/gefgu/en-agentsociety.git $ cd en-agentsociety $ pip install -e packages/en-agentsociety $ en-agentsociety check -c config.yaml $ en-agentsociety run -c config.yaml $ en-agentsociety ui -c config.yaml