Every base can be run by a person, by a state machine, or by a
language model. All three speak the same order language and are
checked by the same rules, so nothing gets a privileged path and you
can read exactly what an agent decided to do.
Because a model takes seconds to answer, nobody is clicking. Orders
are standing: you give objectives and constraints, and
they keep running until replaced. That is not a workaround for slow
models. It is how you would actually command a fleet you cannot reach
quickly, which is where this is going.
The state machine opponent is deliberately good. A pushover would make
any model look like a strategist, and the question worth answering is
whether it really plans better than a few hundred lines of
if-statements.