The other side of the table

Agents at Work — CC BY 4.0

Every lesson so far has stood on your side of the table — your agent, your scope, your test, your liability. This last one crosses to the other side, because there is always someone there: the applicant whose CV your agent read, the customer whose complaint it triaged, the member whose details it moved. They didn’t choose your tool. They mostly don’t know it ran. And they carry the consequences of what it decided.

This is Anchor 4 — product that serves humanity — brought right down to one person. The test the whole course has been building toward isn’t “can the agent do this?” It’s “does this leave the person on the other side better off, or just leave me faster?”

What you owe them

You don’t need a law degree to name the duties; most of them are just decency made specific by the fact that a machine is now involved.

• That a person, not only a machine, stands behind a decision that affects them. This is the human gate seen from their side. When your agent’s output changes someone’s life — hired or not, approved or not — they’re owed a real human judgment, not a rubber-stamped machine call. You already knew this protects you (Tier 2). It also protects them, and that’s the actual point.
• That you didn’t feed their information somewhere careless. They trusted you with it — often without choosing to. Pasting it into a tool that might store or train on it, sending it offshore without thought, letting an agent hoard it — those are breaches of that trust before they’re ever breaches of a principle.
• Honesty that it happened. If an AI-assisted process touched their application or their case, that’s not a secret to keep. Transparency — that AI was used, against what criteria, and that they can ask for a human to look again — treats them as a party to the thing, not an object it was done to.
• A door back to a human. The single most important one. A person on the receiving end of an automated decision should be able to reach an actual human who can hear “I think this got me wrong” and do something about it. An agent with no route back to a person is a wall, and walls are where unfairness hides.

Why this is the whole course, not a footnote

Go back through the tiers and notice they all resolve here. Scope and least privilege — so the agent can’t harm someone through reach it never needed. Criteria not vibes — so a decision about a person can be explained to them. Identity-blind and adverse-impact testing — so it doesn’t quietly filter them out for who they are. The law — because their rights are the thing it protects. The sovereign option — because their data deserves custody. Every discipline in this course is, in the end, a duty to the person who never sat down at your keyboard.

Build agents that pass that test and you’re doing the thing the whole trilogy was for: not automating people out of the picture, but keeping them — you, and the person on the other side — firmly in it.

The one-pager — for the other side

Because everyone is, at some point, the person on the other side of someone else’s agent, the course includes a standalone page written for them: “When an AI reads your CV.” What to strip before you submit, how to format so a machine reads you fairly, the three questions to ask an employer, and your actual rights. Read it as a deployer — it’s the sharpest possible statement of what you owe — and share it as a citizen. It’s the course’s koha to the other side of the table.

Next

The capstone: put it all together into one deployment brief for an agent you’d actually build — and answer, on paper, the question the whole course asked. Who answers for it?