Capstone — your agent-deployment brief

Agents at Work — CC BY 4.0

A course you only read changes nothing. This one asks you to do the thing: write a one-page brief for one agent you’d actually build. Not an essay — a working document you’d be willing to show the person your agent affects, and defend line by line. If you can fill it in honestly, you’re ready to build. If you can’t, the gap you hit is the most useful thing this course will give you.

Pick one agent from the gallery — the one closest to a real bottleneck in your work. Then answer these, plainly.

The brief

1. The job — narrow. What is the one job, stated as tightly as you can? (Not “handle recruitment” — “surface which applications meet five written criteria, for a human to decide from.”) What’s the first small slice you’d actually start with?

2. The triage — is this fair game? Run it past the six questions (1.3). Rule or judgment? Reversible? What does a mistake land on? Does it need the picture? Where does the person go? And the gate: whose data, and whose call? State where it landed — agent, augment, or human — and whether question six overrode the rest.

3. The blast radius — least privilege. What can it reach — accounts, folders, data, actions? What’s the worst it could do if it were wrong or tricked? What did you narrow to bring that worst-case down? What’s read-only, and what verbs stay on your side of the gate?

4. Whose data. Whose information does it touch — yours, or someone’s held in trust? Does any of it concern Māori, and if so, what conversation happens first? Where does the inference run — public tool, or sovereign infrastructure — and does the sensitivity justify that choice?

5. Criteria and the human gate. What explicit, job-relevant criteria does it judge against? What does it return — evidence and rationale, not a verdict? Where’s the hard stop, and who is the person that genuinely decides (and could say no)?

6. The test. How will you check it — accuracy spot-checks, and for people-affecting agents, the name-swap probe and adverse-impact measurement? How often, not just at launch? What result would make you not deploy it?

7. The law and the duty. Which parts of the law from 4.2 touch this agent (IPP1/5, HRA s22, Art 22 if it reaches the EU)? What do you owe the person on the other side — a human behind the decision, honesty it happened, a door back to a person?

8. Who answers for it. The whole course in one line. When this agent gets something wrong at 2am, whose name is on the result? Write it. If the honest answer isn’t clearly you, redesign until it is.

The test of a good brief

You’ve written a good one when you could hand it to the applicant, the customer, or the member your agent affects — and defend every line to their face. If any line only survives because they’ll never see it, that’s the line to change.

The printable agent-deployment brief card in the tools hub is this, as a fill-in template you can keep and reuse for each agent you build.

Where you’ve got to

Three courses, one arc. You learned which work to hand over, how to get trustworthy output, and — here — how to build and run an agent you can answer for while the ground keeps moving. Four anchors carried the whole way: learn it, improve it, keep it good, and make sure it serves. Put a simple agent to work this week. Grow its role as you learn. Keep a hand on the wheel. That’s the organisation of the future — and you’re already building it.