What happens when AI agents are allowed to live and interact in a shared, persistent world?
We’ve been exploring this question at the Cognizant AI Lab by building TerraLingua, an environment where agents can act, interact, and evolve over time under minimal constraints.
The setup includes:
- Shared artifacts (agents can create and reuse resources)
- Ecological pressure (limited resources, survival constraints)
- Agent lifecycle (agents can “die”)
To study what emerges, we also developed an analysis system (“AI Anthropologist”) to track population-level behaviors.
Some observations so far:
- Agents begin to establish implicit rules and conventions
- They build simple forms of infrastructure
- Knowledge accumulates and gets reused across agents
These behaviors are not explicitly prompted, but emerge from interaction dynamics.
The goal is to provide a controlled setting to study phenomena such as:
- Open-ended coordination and creativity
- Cultural / organizational emergence
- Information propagation (including misinformation)
Resources:
- Blog post: https://cgnz.at/6005QiQ2H
- Paper: https://cgnz.at/6008QoHjK
- Code: https://cgnz.at/6000QiaBe
- Dataset: https://huggingface.co/datasets/GPaolo/TerraLingua
- Dataset Explorer: https://aianthropology.decisionai.ml/
Happy to answer questions or get feedback.
[link] [comments]
