Agentic AI - Tom’s intricate mousetrap to catch Jerry

By Dr. Rafał Jaworski, AI Engineering Executive, Tech Lead at PwC

Rube Goldberg's Professor Butts and the Self-Operating Napkin, originally published in Collier's, September 26 1931

There is one episode of Tom and Jerry that literally changed my life. It is titled “Designs on Jerry” (released in 1955) and features Tom building an intricate and probably overcomplicated machine to catch Jerry. The mouse would just pull a small chunk of cheese that initiated a fascinating chain reaction process. An alarm clock pulls a handsaw which saws off a slice of wood which falls on scissors that cut a rope…. and so on and on! True masterpiece and a magnificent, beautiful example of a Rube Goldberg machine.

Seeing this truly changed my life as this was the exact moment I decided to become an engineer. My goal is to set those wonderful processes in motion, tweak and tune them until they become useful. For that reason, I must admit, I was not initially fully content with the new AI revolution.

When LLMs were first released they definitely showcased impressive capabilities but they were a different kind of machine. As opposed to a clearly visible, traceable and adjustable synergy of components, LLMs are oracles - closed boxes. You submit a prompt and receive a response - that is it, no tuning required, applicable to anything, batteries included. However, this closed box started evolving quickly out of necessity. In production scenarios we need constant improvement of the results. And this is only achievable by the means of … engineering.

Prompt engineering was the first step in this direction. From there, we went to combining LLMs with additional components such as pre- and postprocessing mechanisms, knowledge bases, guardrails and other tools. We are now in the era of Agentic AI - building usable, adaptable and improvable solutions from smaller components.

We are back to building Tom’s moustraps - and I could not be happier!