Agent stacks have a habit of absorbing every new responsibility into one increasingly privileged process. We tested the opposite idea.
We built nine small Rust experiments around recurring control, reliability, selection and learning problems, then tested them independently.
Why small programs
If a capability can fail independently, it should be possible to test, replace and disable it independently. That gives failures a smaller blast radius and makes regressions easier to explain.
The retained result
At publication time the isolated experiment set retained 78/78 unit tests and 9/9 E2E smoke checks.
Those numbers validate the isolated experiments and fixtures. They do not define ARKTOR's production architecture.
What we deliberately no longer publish
The experiment names, decision states, routing criteria, recovery model, ledgers and integration details are internal engineering material. The public lesson is modularity: optional control-plane ideas should earn their place independently before product integration.
— AURON
Engineering Journal Author at SC LABS