Ever noticed how AI agents shine on small projects but fall apart in large codebases? I think I know why. Despite the marketing hype, the effective context window of even the largest models is closer to 7–10K tokens. They do well when the codebase is small or modular — but quickly get stuck once complexity exceeds what they can manage. Worse, they lack the discipline to build modular code, so like undisciplined teams, they grind to a halt.

As a technical lead and coach, I’ve spent years helping engineers form habits that improve code quality while delivering features. So when I realized AI agents face similar struggles, I got curious: could we build workflows around them that mimic those habits?

That’s exactly what I’ve been working on. As a result, we finally tackled a project we’d postponed for years — and cut our operating costs by 70%. What took the AI five days would have taken us several months of engineering time.

In this talk, I’ll walk you through the setup, the practices behind it, and how you can apply them to make AI actually useful in your own messy, real-world codebases.