Part 2: agents need infrastructure, not just models
A serving split that halves response time without touching the model, and a firewall that keeps real credentials out of the sandbox entirely.
How we build
Splitting attention from experts in the serving layer
Attention layers are bound by memory bandwidth and expert layers by compute, so running both on the same accelerators wastes capacity on whichever workload loses. Split into two services scaled on their own, response time on one model falls by 47 per cent.
We tell a client that the same workload can cost half as much again purely through where it has been placed, and that none of this is visible from the application above it. A serving topology chosen a year ago is a standing cost decision nobody is revisiting.
How we assure
A credential firewall for agent sandboxes
The sandbox holds opaque proxy tokens. All outbound traffic runs through a daemon on the host, which swaps the token for the real credential before the call goes upstream. The real key never crosses the container wall.
Our position is that one prompt injection, one compromised dependency or one rogue tool call can dump the environment out of a sandbox in seconds. That is the whole attack. This closes the gap in the infrastructure rather than trusting the model to behave, which is the only version of the control that survives contact with an adversary.