A six-part architecture series with one governing principle: put the model where being wrong is cheap and recoverable, and put code where being wrong is expensive or unauditable. Most agent designs fail because they apply it in reverse.

1. Put the Model Where Being Wrong Is Cheap — The governing principle, and the line nothing may cross upward.
2. Use the Model to Write the Parser — Use the model once per class of problem, not once per instance.
3. Where the Model Is Genuinely Load-Bearing — The short list of jobs only a model can do.
4. An Agent Is a State Machine, Not a Loop — Structure the harness so every run is replayable.
5. The Model Upgrade That Quietly Breaks You — Behavioural and adversarial evals, because the failure is silent.
6. The Model Is the Least Defensible Part of Your Agent — Anyone can call the same API. The moat is the artifacts.
7. Does a Smarter Model Change Any of This? — The companion question — and why the answer is mostly no.
8. What a Good Eval Set Actually Looks Like — Curated cases that discriminate — and a grader that can’t grade itself.

← All series