The work
A convincing agent demonstration can hide how the result was obtained. For practitioners, the more useful questions are what knowledge the agent had, what it was allowed to do and how its output was checked.
My contribution
I designed and facilitated a three-hour session at Prosus AI House in Amsterdam, with MLOps Community. The opening compared coding agents on the same task with different knowledge and context. Participants then worked with agent configuration, practical challenges and evaluation criteria.
The opening task was to build a simulation of artificial life with a plot tracking its total mass. The plot revealed changes hidden by the moving shapes. Participants could then check whether an output met explicit criteria, including whether the diagnostic kept updating.
What was delivered
A keynote, practical workshop and reusable teaching materials.