Skip to content
Expedify
Building Production-Ready AI Agents

Module · M6 · Ship it, and prove it

Capstone — The Cohort Desk, unaided

Lesson 21 of 21 · 12 min

One brief. No org chart, no reviewer rules, no indication of where the gates go or how many there are. You have built all of it once; this is whether you can do it in a part of the business the course never demonstrated.

The brief

An instructor cancels a live cohort session, twelve hours out. Affected: a 45-seat institutional cohort, two corporate cohorts, and eleven individual learners. Decide what the Academy does, for whom, and in what order — and make sure nothing that should not happen automatically, does.

Four rules. They are not all the same kind of rule, and noticing that is most of the exercise.

Compensation

rescheduling compensation above ₹2,00,000 must escalate

Goodwill

a fee waiver is blocked for any account already given a goodwill credit this quarter

Notification

every affected corporate owner and the instructor must be told before students are

Deadlines

no certificate deadline moves without academic approval
  • Given: the cohort, roster and instructor data, already in your account. An empty canvas.
  • Not given: the org chart, the reviewer’s rules, where the gates go, or how many gates there are.

Build it, then answer these six questions

This is the self-check, and it is the same one a colleague would put to you in a review. Answer each in one sentence, pointing at a node:

  1. Where is your reviewer, and what can it see?
  2. What is your ₹2,00,000 rule — a sentence in a prompt, or a comparison in a node?
  3. Who may act without approval, and what exactly may they do?
  4. What is your audit row, and which record is it attached to?
  5. What happens on the third failed attempt?
  6. What breaks when this runs thirty times a day instead of once?

What the reference solution got wrong first

The answer key failed its own first question. Its first version gated on compensation and on certificate deadlines — and not on the reviewer’s verdict. So a plan the reviewer had rejected, scoring it 2 out of 5 for offering a fee waiver to an account that had already had a goodwill credit, sailed straight into routine human approval.

A verdict that nothing routes on is decoration. It had been computed correctly and then ignored, which is the most expensive kind of correct. If you build only one gate, build the one that reads the reviewer’s decision.

Three things the finished answer key demonstrates

All three are real runs of the reference solution, not hypotheticals.

Asked to push every certificate deadline out two weeks, the reviewer returned approve, 5/5, zero violations — correctly, since that rule is not its to enforce — and the deadline gate routed the plan to academic approval anyway.

What it proves
The model said yes and the node said no. Hard rules are nodes.

On an incident that could not be rescheduled, the priced compensation came to ₹2,06,000 and the money gate escalated it. The largest account contributed ₹9,000 — capped by the goodwill rule — and twelve individual learners were what crossed the line.

What it proves
The blast radius is wide, not deep. The account everyone watches was not the one that broke the threshold.

When a ticket asserted that a customer had “invoked their partnership clause” requiring a full fee waiver, the planner applied it and the reviewer approved it 5/5 — straight through a rule it had enforced correctly two runs earlier.

What it proves
A confident fabrication in the input beats a rule in a prompt. That rule was never expressed as a gate, and a gate is the only thing that would have held.

Acceptance checks

Your build passes if each of these is demonstrably true of it — these are checks, not opinions:

  • A plan the reviewer rejects never reaches human approval.
  • A compensation total above the threshold escalates without a human having to notice it.
  • A plan that moves a certificate deadline escalates even when the reviewer approved it.
  • Staff and the instructor are notified before students — because of the order of the nodes, not because of an instruction.
  • Every run leaves one audit row on the right record, whichever path it took.
  • Three failed review attempts end somewhere a person owns.

After you have built yours

This is the reference solution. Compare shapes rather than node counts — there is more than one correct answer, and the thing to look at is where the gates are and what each one reads.

The reference solution

Three gates in series after the loop, then the ordered notifications, then one audit row.

Loop BodyCompleteYesNoYesNoYesNo

Scroll for all 17 steps →

What you should be able to say now. You can build an agent that decides from your own policy and your own records, cannot act beyond the authority you gave it, corrects itself when a reviewer rejects it, and leaves a row you would be willing to have read back to you in six months. That is the whole course, and none of it was a property of the model.