Why every new AI agent needs a short course, not a memo

A memo tells people an agent exists. A short course makes sure they can work with it: what it may do at each depth, why it holds or refuses, which sources it trusts, and what to do when its certificate lapses. Every new agent changes what people approve, so each deserves an hour of assessed training, not an email.

What does a memo actually achieve?

The launch memo is a familiar artefact. It names the agent, describes its job in a paragraph, lists a contact for questions and links to a user guide. Some people read it closely, most skim it, and a few never open it.

That would be fine if the agent changed nothing about anyone's decisions. It does. From the day it goes live, people start approving its drafts, releasing its held actions and explaining its refusals to customers. The memo has told them none of the things they need to do that well, and it has no way of finding out who understood it. When an approval later goes wrong, "we sent a memo" is a weak answer to the question of whether the reviewer was prepared. Memos also age badly: the agent changes at its next certificate renewal, and the memo stays as it was.

Why is each agent different enough to need its own course?

Because the shared vocabulary only gets people part of the way. Everyone who works with agents should know depth, the four verdicts, provenance and lapse. That is taught once.

Each agent then has specifics that cannot be guessed. Its role scope: what it is allowed to touch. Its depths per capability: perhaps it executes routine matches at depth 3 but only drafts payments at depth 2. Its sources and their classes. The business's own limits that cause it to hold. The red lines that reviewers must never approve. And its likely failure modes: the cases where its drafts tend to be weakest. A reviewer who knows the vocabulary but not these specifics will misread the agent's verdicts in exactly the cases that matter.

What belongs in the short course?

Keep it to about an hour, built from the agent's own material:

1. The job and the depths. What the agent does, capability by capability, and which actions reach a person.
2. Why it holds. The three or four most common reasons, each shown on a real case.
3. Refusals and withheld actions. What it refuses and why, and what is withheld at the platform floor for every agent.
4. Sources. Which facts it treats as sealed, signed, declared or external, and what that means for approval.
5. Abnormal states. What to expect when its certificate lapses after 90 days, or when a run that clears no depth drops it to depth 1 the same day.
6. A short test on real cases. Five to ten held actions, decided and marked.

The test is the part a memo cannot replace. It tells the business who is ready to approve and who needs more practice, and it creates a record that the training happened and worked. A business can adopt the practice of granting approval rights for a new agent only to people who have passed its course.

Where does the material come from?

From the agent's shadow period and its first weeks. Every hold, edit and refusal is a potential case. Refusals are especially useful, because each one shows where the business has drawn a line and why; every refusal is a lesson for the team, not just the agent makes that case in full. Collect them from the start, and the course largely writes itself.

Not everyone needs the same depth. A learning needs survey across the teams the agent touches shows who will approve its actions, who only needs to explain them, and who will not meet it at all. Sector matters too: who in a utility needs AI agent training first is a different list from who in a private clinic needs AI agent training first.

What is the trade-off?

Course sprawl. A business running a dozen agents could end up with a dozen courses, each needing updates whenever an agent changes, and staff could tire of being sent on another one.

Three habits keep it manageable. Teach the shared vocabulary once, so each agent's course covers only what is specific to it. Use one fixed structure for every agent course, so writing a new one is a matter of filling in sections. And manage them like any course catalogue: each course has an owner, a version tied to the agent's current certificate, and a retirement date when the agent is decommissioned. A short course that is current beats a long one that describes last quarter's agent.

There is a smaller trade-off in time: an hour per approver per agent. Compare it with the time spent unpicking one wrongly approved action.

What should you do differently?

Stop launching agents by memo. Make a short, assessed course part of every agent's go-live checklist, owned by the agent's owner and built from its shadow-period cases. Gate approval rights on passing it. Update it when the agent gains a capability or depth, and retire it with the agent. The broader sequence for training people around agents is in the guide to training teams for AI agents.