Field Note
Agents Don't Replace Your Rules. They Wrap Them.
The moment you let your people build with AI, the fear kicks in: the rules go out the window. Compliance, sign-off, the way things have to be done here — all of it loosens the second a non-expert can build their own tool.
Done wrong, that fear is correct. Done right, it’s backwards — the rules get more reliably enforced, not less. The difference is which way round you build it.
The common move is to take the software you already have and bolt AI onto it — a chatbot in the corner, a “summarize” button. Pete Koomen, a YC partner, has the best name for this: the horseless carriage. Early cars looked like carriages with the horse removed, because nobody had rethought the vehicle yet. Most “AI features” are the same — an engine strapped to the old shape.
The inversion is the whole game. Don’t put AI inside your software. Put your software inside the AI.
Instead of a rigid app with a little AI slotted in, you run an agent that’s allowed to use a set of deterministic tools — your rules, encoded, doing exactly the same thing every time. The agent brings the judgment about what to do; the tools carry the parts that must not vary. It’s the difference between hoping someone follows the checklist and building the checklist so it runs itself.
For anyone in regulated work, underline this part. Those wrapped tools — call them skills — are exactly where your traceability and sign-off live. The rule stops depending on a tired human remembering it at 5pm on a Friday. It’s encoded once and invoked every time, by the newest person and the most senior, identically.
I build this way at my own bench. When I close out a piece of work, an agent runs a skill I wrote that won’t let it be called done until it’s been tested, recorded, and finished in order — and won’t sign off on anything complex without showing me exactly what changed first. The agent can’t skip the check. The check is the skill. That’s not a limit on the agent; it’s how I trust it with more.
That’s empowerment your compliance team can actually live with. Not blind trust. Not a lockdown that kills the whole opportunity. Doctrine — the right way, written once, wrapped so it runs.
This is one of several threads in a book I’m writing on running AI like a readiness problem, not a purchase.
What’s one rule in your shop that lives only in somebody’s head — the check everyone’s supposed to do and half the team forgets? Start a conversation and name it — that’s the first one worth wrapping into a skill.
Sources: Pete Koomen (Y Combinator), “AI Horseless Carriages” (koomen.dev, 2025) and YC’s “The Lightcone” — the case that most AI apps bolt AI onto old software instead of building the software around the agent.