We Model What Agents Know, Intend and Assume

AI agents are taking over critical workflows, but they fail in predictable ways as interactions gain depth. They lose context, contradict earlier decisions, and break down over long workflows. These are systematic failures with real cost.

We are building a control layer for complex multi-party interactions between humans and agents. Our system operationalises an Artificial Theory of Mind for agentic AI by modelling what participants know, intend, and assume. This allows agents to maintain coherence over time and complete tasks more reliably.

Alignment is often framed as a safety problem. We see it as a state problem. An agent that cannot retain what has been established cannot stay aligned with user intent.

Our approach operates at the post-training level, adapting open models for coordination, role coherence, and context retention. It reduces dependence on proprietary systems and gives organisations direct control over how their agents behave. Our interaction architectures provide standardised structures for reliable agent behaviour across high-stakes domains.

We are validating the approach with early partners and universities while looking for customers to work with and team members to join us.