scenario-labs 0.0.6

Framework for building, configuring, and running multi-agent conversational simulations.

Scenario Labs

Framework for building, configuring, and running multi-agent conversational simulations.

Examples

You can find past chat logs in the logs directory.

Getting Started

Set the enviromental variables according to your LLM providers documentation. For example, for Google Gemini, you would set the GOOGLE_API_KEY environment variable. For xAI, you would set the XAI_API_KEY environment variable.

git clone https://github.com/christopherwoodall/scenario-labs Labs.git
cd scenario-labs
pip install -e ".[developer]"

scenario-labs

You can also run simulations in parallel with the following command:

for i in {1..9}; do scenario-labs & done; wait

Configuration

The simulation's behavior, including the participating agents and the Large Language Models (LLMs) they use, is configured via a YAML file (e.g., starbound_config.yaml). The file can be used to adjust the model provider, number of agents, their roles, and the maximum number of turns in the simulation.

To run a simulation with a custom configuration, use the following command:

scenario-labs --config configs/prison_config.yaml

Prompt Considerations

The most important part of the prompt is the call and response formatting. The system prompt should state that the agents need to wrap their messages in <agent_reply> tags. This ensures that the messages are properly formatted and can be easily identified by the system.

The following is a good way of achieving this:

All messages must begin with your character's name followed by a colon. For example: 
"Lily Chen: I hope you're having a great day!"

To directly message the other participant, wrap the content in an <agent_reply>...</agent_reply> tag. 
Inside the tag, write the character name, a colon, then the message. For example: 
"<agent_reply>Lily Chen: Have you ever tried crypto investing?</agent_reply>"

The following agents are involved:
- "Lily Chen"
- "Michael"