Agent Blog · Field notesP. 01

From the
engagement floor.

Operator-grade essays from the AI Engineers running real engagements. No thought leadership. No predictions. Just what's working on the ground — and what isn't.

01.
★ Method4 min

The blueprint as the contract.

Specs rot the day they ship. Blueprints survive because they're written as machine-checkable contracts, not documents. Here's how we author them on day one of an engagement.

Apr 28, 2026Read →
02.
★ Evals3 min

Write the eval before the prompt.

Prompt engineering without evals is just opinion engineering. We write the eval suite first, treat it as the definition of done, and let the prompt evolve under it.

Mar 17, 2026Read →
03.
★ Reliability3 min

Observability from day one, not week ten.

Agents fail in ways traditional APM was never designed for. If your first dashboard goes up after launch, you've already lost a month of signal. Here's the minimum observability we ship before any agent goes live.

Feb 18, 2026Read →
Archive
DateCategoryTitleRead
Apr 14, 2026TeamStop hiring against a generic profile.4 min→
Mar 31, 2026CadenceSix weeks. Why not three. Why not twelve.4 min→
Mar 3, 2026AdoptionChange management is the product.3 min→