research // what we test, what we learn
We run our agents on real boxes, so we test on real boxes too. Here's what we measure, how we measure it, and what we change because of it.
benchmark
We gave four coding harnesses the same model, the same tasks and the same 1 GB box our agents live on, and scored them on hidden tests. With a strong model they all got everything right. The difference was the bill.
Read
launch notes
Six choices that shape how an agent behaves, a box of its own, your own model key, SSH into its brain, and a button that lets your own AI do the setup.
Read