From factory sensors to AI agents

I started my career with factory sensors: real devices, on a real production floor, sending real data over an industrial network. A sensor is unforgiving. It either reports the right value on time, or it doesn't, and everything downstream breaks.

That discipline carried over when I moved to AI-agent systems. An agent that calls tools and makes decisions has the same failure mode as a sensor with a flaky connection: it looks fine until the one input it didn't expect shows up.

• A sensor system taught me to distrust a value until I checked its source.
• An agent system taught me the same lesson, at a higher level: check a tool's real output, not the response I assumed it would return.

I have carried this from SmartPlant and SmartChronos, through Société Générale, to the startups I now help build on AWS. The tools changed. The discipline of not trusting an unchecked value didn't.