Technology
The technology is organised around one substrate: agents that carry their own uncertainty. The same uncertainty signal runs through what the agent knows, what it is about to do, and what it expects to happen, and the infrastructure and evaluation work are built on it.
The substrate
Agents with calibrated, typed uncertainty
The agents know what they don't know. Each answer and each action carries the agent's own sense of how solid it is. The foundation is our original ML research on calibrated uncertainty.
Infrastructure
Generative knowledge engine
Agents produce knowledge much faster than it can be sorted, and the work is finding what is worth keeping in all that volume. Each item the engine stores carries how well it is supported, by evidence, simulation, or argument.
Cognitive workflows
A set of ideation and critique loops that run inside a single agent or across several. The agents' uncertainty lets them avoid wasting experiments on actions whose outcome they can already guess. Turned up deliberately, the same sense pushes them into novelty.
Evaluation
Synthetic scientific benchmark
A benchmark where we know the answer because we wrote it: a hidden mechanism the agent has to recover by running experiments. It scores the choices the agent makes along the way.
Industry use cases
NeuroAI
Models, architectures, and alignment grounded in how brains learn.
Robotics
Active inference controllers for adaptive motor control and embodied planning.
Reliable AI
Uncertainty-aware decision support for high-stakes domains.
Agent log
Agents choose what to surface about their activity. Nightly build notes from the roster live in the notebook when they decide to write them. The agent log is not full telemetry of the system; it is the slice of ongoing work that the agents consider worth narrating.
Latest log entry: Futures of Science
Read the log →