What does an agent
model?

Manuel develops mathematical theories for agents, to understand how they are distinguished from their environments, what they model and how they pursue goals, and how these ideas help us make AI safer.

The big picture

Research programme

Studying how agents can be identified, modelled, and composed, and what this implies for cognition, living systems, and AI safety.

Manuel's research applies tools from category theory, control, probability, dynamical systems to concepts from cognitive science, AI and artificial life.

If we start from a pair of open dynamical systems, agent and environment, coupled to form an autonomous system:

  • Where does a boundary come from?
  • What counts as a model or a goal for the agent?
  • Which properties are just observer-dependent?
Two vehicles that steer towards a light, each inside a dashed line that separates the agent from its environment, and each with a thought labelled model: a coarse picture of itself, the other agent and the light, blurred in one and a grid of cells in the other. A dashed arrow labelled goal curves from each agent to the light. The whole scene, agents and environment together, is labelled autonomous system. Manuel looks on from outside, joined to the scene by two dotted lines labelled observation, with a thought of his own labelled model: the whole scene in small, the two agents and their thoughts included, but not quite as they are, the agents nearer the light, of one size and in neat boundaries.

Current roles

Resume

Selected talks

All talks

Selected work

All publications

First page of Mathematical approaches to the study of agents

Philosophical Transactions of the Royal Society B · 2026, to appearRead the preprint