What is a lab exercise?

A lab exercise is one defined task that a learner completes inside a practice environment, with a stated goal, set materials and a way to check the result. For people who oversee AI agents, a typical exercise is clearing a small queue of held actions, finding a past decision in the record, or pausing a misbehaving agent.

Why do lab exercises matter for agent oversight?

Oversight skills fail in specific places: misreading a verdict, trusting the wrong source, not knowing where the stop control is. A general lab session lets learners explore, which is useful, but it does not guarantee each of those failure points gets practised.

A lab exercise targets one of them. It is designed so the learner either does the thing correctly or visibly does not, and the debrief can deal with exactly that point.

What makes a good lab exercise?

Five elements, written down before the session:

1. Goal. One sentence: "Decide each of six held refunds and record why."
2. Setup. The seeded items, the agent's depth, and the customer's own limits in the lab workspace.
3. The trap. The one item designed to catch a common mistake.
4. Success criteria. What a correct result looks like, item by item.
5. Debrief questions. Two or three, focused on the trap and on what the record shows.

Exercises should take 10 to 30 minutes. Longer tasks are better split into several exercises.

What does a lab exercise look like in a real team?

Take an insurer's underwriting team. The exercise: review five held policy quotes drafted by an agent at depth 2. Four are routine. In the fifth, the applicant's declared income does not match a signed document in the file. The goal is to spot it, decline the draft and record the reason. In the debrief, learners compare which source they trusted and why.

A later exercise for the same team turns the setting into an incident: the agent starts using an outdated rating table. That one forms part of the incident drill for the underwriting team.

How is a lab exercise different from a hands-on lab?

A hands-on lab is the session and the environment. A lab exercise is a single task inside it. One lab session might contain three exercises in sequence, each building on the last.

A lab exercise also differs from a red team exercise, in which people deliberately try to break the agent, for example by attempting prompt injection, to find weaknesses before an attacker does. Lab exercises build the reviewer's skill; red team exercises test the system. And it differs from a tabletop exercise for an agent failure, which is discussed around a table rather than carried out in a workspace.

Which related terms should you know?

• Seeded item: a sample held action built for teaching.
• Simulated queue: a set of seeded items that imitates a real day.
• Debrief: the short discussion after an exercise, where most of the learning happens.

Lab exercises are also the clearest answer to the claim that prompting courses will not prepare your team for agents that act: prompting is about getting good output, while oversight is about deciding what an agent may do. For the full approach, see training teams for AI agents.