Islam Magdy

In April I ran a full-day, hands-on CrewAI workshop for about 21 engineers across the mortgage domain's teams at Nationale-Nederlanden. The goal was not to teach a framework — frameworks change — but to give the domain's engineers a working mental model of agents as running systems: roles, tools, hand-offs, and the failure modes when a step silently does nothing.

What landed

• Building their own agent in the first hour. Every participant had a running agent on their own use case before lunch. Ownership beats examples.
• Failure demos. Showing an agent that reports success while having done nothing taught more than any slide. Regulated teams already distrust silent failure — naming it in agent terms made the abstraction concrete.
• Cost as a first-class output. When each run prints tokens and dollars, the room stops asking whether agents are magic and starts asking what a run should be allowed to cost.

What did not

• Framework-first teaching. The CrewAI-specific abstractions evaporated within a week; the lifecycle thinking did not. Less framework would have cost nothing.
• One shared use case. The teams' day jobs differ enough that the joint exercise fit nobody. Individual use cases, then a shared retrospective, worked better.

The follow-up is the part that matters: a workshop changes vocabulary, not systems. The teams that built something of their own in the room are the ones still using it.