Applied Research

A language model is not a whole brain.

Models make agents capable. Reliability comes from bringing the right kind of thinking to each part of a problem, and a language model is one part of that. We are building the rest, and we call it Cortex.

Plate I. A search in progress. Most branches are pruned; the line it currently prefers is marked, and moves when a better one appears.

The problem

Reliability is a property of the architecture.

Better models help, and still the errors compound over a long task and nothing tells you when an answer is wrong. In work that moves money, commits resources or runs unattended for hours, that is the difference between an agent you can hand the job to and one you have to watch.

That work demands determinism wherever determinism is available. It demands reasoning at the right level of abstraction rather than the lowest one. And it demands a trace that can still answer why a decision was made, long after the moment has passed.

Cortex beat the frontier models on τ³-bench banking, a hard test of knowledge work under policy, and we did not tune prompts or fine-tune anything to get there.

Cortex on τ³-bench

Cortex

Not everything should be a language model.

We spent decades building software that was precise, with algorithms of known bounds and recipes proven and reused a million times over. Reliability comes down to taste in picking which of them each part of a problem needs, and knowing what to optimize. In Cortex the cognitive core directs the work. The language model does the listening, the talking and the semantic judgement, and other kinds of compute do the rest.

Abstractions are not dead at all. We want agents built the way compilers are, as bounded languages stacked one on another, each compiled into the one below, with a language model writing only the top layer.

Figure 1 We want the cognitive core to direct the compute. As a problem is worked, the compute moves from one kind to the next, and whatever gets settled is kept for the next problem.

Recipes, abstractions and taste

Working notes

What we believe.

N-01 Hallucinating intelligenceLanguage models are trained to write what sounds right. Most of the time that is also what is true, which is exactly why their mistakes are so hard to catch. We sort the questions an agent faces by the kind of compute that should settle each one, and list what would prove us wrong. Diagnosis · October 2026
Principled Machines · 4 min N-02 Recipes, abstractions and tasteComputer science already has most of the recipes reliable agents need. We look at program synthesis, small languages stacked and compiled down, process calculi and planning that adapts, and argue the skill is in picking among them. Approach · October 2026
Principled Machines · 5 min N-03 Cortex on τ³-benchCortex beat every published frontier model on τ³-bench banking, by about as much as eight months of model releases had added. We describe what we did, what we deliberately did not do, and why a specialized system should win this kind of work. Benchmarks · October 2026
Principled Machines · 3 min

All notes

Team

Built by senior operators from

Autonomy matters when an agent can own the whole job.

If people have to keep redirecting and repairing an agent, they still own the work. Cortex is in private early access. We are hiring AI scientists and AI engineers.