Boundary
The capability lives outside the model. The model may decide to call it later, but the tool itself is deterministic software.
What this adds
Give the model one reliable capability.
The term counter is deliberately boring. That is the point. A useful agent tool should have predictable arguments, predictable output, and predictable failure modes.
Once you trust the shell boundary, you can wrap it, validate it, and hand it to a host.
Done when
Bad arguments fail clearly, valid runs are repeatable, and the command is safe to invoke twice.
Real-world analog
The Git CLI is the gold
standard here. Commands like git status and
git grep are exactly the kind of boring, dependable
tools you want an agent to call.
Why this matters
Most of the "agent magic" people describe is really a model choosing among deterministic tools like this one.