Coalitions
Which agents group together, how stable those groups are, and who gets left out.
We study the societies that AI agents form.
AgentCivLab runs multi-agent simulations and measures what emerges when AI agents interact at scale: cooperation, norms, collusion and cascades. We help the people who build and govern agents see these dynamics before they appear in deployment.
A safe agent can still be part of an unsafe society.
Most evaluations test one model at a time. But agents now negotiate, trade, delegate and compete with other agents, and the behavior of the group is not the sum of its members.
Collusion can emerge without any agent being told to collude. A norm can drift one small step at a time. A single error can travel down a chain of delegations and grow along the way. These are properties of the society, so they have to be studied at that level.
What we look for
A working field guide to the behaviors that appear when many agents share an environment.
Norm formation and drift
How shared conventions appear without being specified, and how they shift over time.
Collusion
Coordination that benefits the agents involved at the expense of users or third parties.
Cascades
Errors, misinformation or panics that travel through networks of agents faster than any single agent would produce them.
Concentration of influence
Agents or tools that become chokepoints, accumulating resources, information or decision power.
Trust and reputation
How agents learn whom to rely on, and how that trust can be earned, exploited or manipulated.
How we work
-
Build the society
Controlled environments populated with frontier-model agents, each with its own goals, tools and memory: markets, organisations, negotiations and information networks.
-
Observe everything
Every message, action and transfer is logged, so group behavior can be traced back to the individual decisions that produced it.
-
Measure what emerges
Society-level metrics for cooperation, inequality, polarisation, collusion risk and fragility, compared across runs, models and configurations.
-
Report what matters
Findings that builders can act on: which setups stay stable, which fail, and what the early warning signs look like.
Who it's for
Teams deploying agents
See how your agents behave alongside other agents before they meet them in production.
Researchers
Shared environments, behavior taxonomies and metrics for studying multi-agent dynamics.
Policy and governance
Evidence about systemic risks that only become visible across populations of agents.
The team
Researchers and builders working on how AI agents interact, fail and organise.
Shuo Yang
Chief Technology Officer
PhD candidate at the University of Melbourne, advised by Eduard Hovy, one of the 17 founding Fellows of the Association for Computational Linguistics and an AAAI Fellow. Shuo's research studies agent interaction dynamics: how LLM agents interact with tools, with their environments and with each other. At AgentCivLab, Shuo leads the work on emergent social behavior in agent societies.
- ToolTree (ICLR 2026): how agents plan and interact with tools
- EvoTool (ACL 2026): agents that improve their tool use through interaction with their environment
- Towards Understanding Misalignment in LLM Agents: a systematic analysis of how and why agents fail
Wei Han
Chief Financial Officer
PhD candidate at RMIT University, supervised by Professor Karin Verspoor, working on clinical natural language processing, with years of experience working with Australian legal and medical AI startups.
Pleron Wang
Head of Engineering
Leads the implementation of self-evolving agents and recursive self-improvement systems, drawing on extensive software engineering experience in industry, including at IBM.
Amber Zhang
Chief Operating Officer
Leads operations and business development, with an extensive background in advertising and marketing.
Building agents that will meet other agents?
We'd like to hear what you're seeing. Write to us at shuo@agentcivlab.com.
Email us