The stupid part is built in. Prediction is not truth, the graders' taste is trained into it, and its memory runs out at the edge of the desk. The next model will be more capable. Do not wait for it to be reliable, because reliability is what you build.
Almost everything an agent does well comes from what you decide it reads: one clean page instead of the restaurant's whole history, nothing stale, every piece weighed the same, chosen by somebody who knows the business. And every mistake it repeats is a line you have not written yet, in the rules file, in a skill, or in a gate.
Proof never comes from the agent's word. It comes from a gate that can refuse, a regression suite that remembers yesterday, and a log of every session, read in batches by another agent that quotes its evidence.
The student who asked about one main agent over all the others was asking the right question a little early. Build one agent. Make it work really well. Then build the second. Nobody has two years of experience with this yet, and the people who start now are the ones who will.