AI systems, context, and execution.

Most AI work fails. Not because the tools are wrong — because there is no operating model around them.

It's a systems problem.

Field notes for engineering leaders building AI-native operating models.

What changed

Generation is cheap. Carrying is not.

AI made code easy to write and harder to maintain. The economics inverted; engineering practice has not.

The harness →

Adoption is visible. Operating model is not.

Tools register on dashboards; the things that make AI work or fail register only in retrospect. The visible variable is the wrong one.

Vibe vs. reliable →

Code is no longer the artifact. The specification is.

When agents author code, the spec becomes the durable thing. What gets written is what gets versioned and re-run.

Specify · direct · validate →

Execution moved. Judgment did not.

Agents type; engineers decide. The labor shifted; the org chart has not.

Architect-CEO →