Safe multi-agent communication
Why
Soon most work will be done by swarms of agents cooperating and competing with each other, sometimes with differing ownership. However, today’s models were trained only to be good one-on-one assistants. Put many of them together and they collude, reward hack, converge on the same answers, and experience correlated failures.
Models have not yet learned during training how to communicate safely and effectively with each other. We aim to address this problem.
Work
-
Benchmarks
We need measurements that describe how good agents are at communicating with each other, and we need to measure levels of risk we are incurring from social behavior.
-
Environments
Environments for groups of agents.
-
Post-training
Our north star is giving agents the ability to communicate with each other safely and effectively.
Track record
We co-authored Agents of Chaos (Shapira et al., 2026), the first published red-teaming study of interacting agents, covered by Science and Wired. After that, we spent months doing follow-up projects with a frontier lab.
Hiring
We’re hiring a founding engineer and strong researchers who are interested on multi-agent RL. Email contact@agents-of-chaos.ai.