Journal
Notes from the lab.
Research notes, field guides and essays from our work. We publish when something has held up in practice.
Research note · 16 Oct 2026
The least expensive model that qualifies
Routing is two questions asked in order: which models can do this task, and which of those costs least. Most routing mistakes come from asking them the other way round.
6 min readField guide · 13 Oct 2026
Evaluate on your own tasks
Public benchmarks tell you how a model does on someone else's work. A small suite built from your own tasks tells you what will happen when you change models.
7 min readEssay · 9 Oct 2026
Agents need a home base
An agent with no shared memory starts every task as a stranger. A home base gives agents and people the same board, the same notes and the same record of what was decided.
6 min readPrinciple · 6 Oct 2026
Keep people in charge of the machine
Agents can run unattended. That is different from running unsupervised. The difference is a handful of deliberate places where a person decides.
5 min read