How software teams keep control of codebases that AI now writes most of.

I'm a hands-on architect and engineering leader. Sixteen years across data systems, machine learning, and full-stack work, most recently as Director of Engineering at a US digital-health startup, where my teams shipped AI-generated code into a HIPAA and SOC2 production stack every day. Before that, eight years running my own consulting practice.

Agents write the code now. The work has moved from writing code to reading code, and it lands on the senior engineers: either bogged down in the review queue, or rubber-stamping PRs they have not read. Your team has the rules written down somewhere, and the model is probabilistic, so those rules hold only some of the time. Two or three months in, the newest modules are dark boxes nobody can explain. I build the mechanical gates that keep AI-assisted development under control, and I write about what actually works.