The second use case costs as much as the first

The first agent looks expensive because you are inventing everything. The second should be cheaper. In most of the organisations I sit with, it is not. It costs the same, sometimes more, because nothing from the first one can be reused except the slide.

They already have layer one — Claude, Copilot, ChatGPT, Gemini — and layer two, an orchestration path that can run a job when something happens. Almost none have layer three: approval, a record, and a way to reverse the write. Each team invents a ceremony, or skips one. Audit asks the same question either way.

What “reusable” actually is

Not a shared prompt. Not a shared vendor. A written constraint: what may be drafted, what may be executed, what the approver sees, who signs, and how you undo it the same day. If that document is specific to one team’s tooling, it is not a constraint. It is a runbook.

I keep a gate file for this reason. You can take it. Change the names. If you cannot fill it in, do not build. The missing piece on the second use case is almost never another model.

What I got wrong when I had to build it

I let the first hire’s approval live too close to the first hire’s interface. It worked. It did not transfer. The next desk asked for the same invention again, with different nouns. That is how you spend a year and have one agent.

Bind the gate to the write, not to the product surface. Then the second use case has something to inherit besides a demo recording.