We build models and infrastructure that simulate human organizations, helping teams of AI agents work together on hard problems.
Founded by two former engineers at
With experience training large language models.
Our first paper explores how specialized AI agents share evidence, question one another’s assumptions, and revise their conclusions.
The system uses shared memory and independent review to make its reasoning easier to examine.
An early case study of how agents review and correct their reasoning. Predictive performance has not yet been established. The full paper is not publicly available.
More agents don’t automatically make a better team. We study how roles, shared knowledge, and constructive disagreement work in human organizations, then simulate these structures and behaviors in AI teams.
Without clear roles and handoffs, agents can repeat work or leave parts of a problem unresolved.
Evidence and decisions can get lost between agents. The team needs a shared understanding of what it knows and what has changed.
Agents can reinforce one another’s errors. They need ways to question assumptions, check evidence, and correct their conclusions.
Research our founders have contributed to is deployed at some of the world’s largest firms, with substantial economic impact.
We’d like to hear from investors, research collaborators, and teams with hard problems to solve.
Get research updates and hear about opportunities to try our models and infrastructure.
