Architects of Intent
Friday afternoon, you merge an AI-assisted change. Saturday morning, your app is down. The model quietly changed a contract your type system never saw.
This book is about making that failure mode boring to catch. Start
with one bounded loop: declare intent, slice context, generate once, and
gate with deterministic PASS/FAIL checks.
Then reuse that shape across teams: map reality, execute bounded loops, and keep the ledger explicit enough that governance stays legible.
“We don’t trust the model. We trust the loop.”
Start with a Mission Object, a bounded diff, and one deterministic
PASS/FAIL gate.
See how slicing, validators, and circuit breakers keep stochastic generation inside a known blast radius.
Reuse the same loop across teams with explicit control surfaces, evidence, and governance.
The Illusion of AI —
vs. The Reality of Results
Why simply buying a “smarter AI” isn’t enough, and how the world’s best companies are turning unpredictable AI into a reliable competitive advantage.
The Problem: AI is a Wild Artist
The “Vibe Coding” Myth
Most people treat AI like a magic vending machine: you put a prompt in, and hope perfect work comes out.
- It hallucinates.
- It’s inconsistent run to run.
- As you go faster, the risk of total failure spikes.
The “Architecture” Reality
Winners don’t just use AI; they build a factory around it. They use strict rules to catch the AI’s mistakes before you ever see them.
- Intent is clearly defined.
- Rules act as a safety net.
- Mistakes are caught and fixed automatically.
The Secret Math of AI Dominance
Everyone is obsessing over buying the biggest, smartest AI model. But the real explosion in productivity comes from a different formula:
A smaller, cheaper AI in a self-correcting loop will ALWAYS beat a genius AI on its first try.
Your New Competitive Moat
AI models get better for everyone. Prompts can be copy-pasted. Talent moves around. But a finely-tuned system that automatically generates, checks, and perfects work? That is a moat your competitors cannot steal.
The “Deterministic Sandwich” Simulator
Wrap an unpredictable generator in deterministic prep + deterministic validation. The loop retries until it’s admissible.
Who This Is For
- Teams shipping AI-assisted changes to production
- Platform and tooling teams building AI-assisted development workflows
- Engineering, security, and compliance leaders establishing AI governance
- Anyone who has been burned by code that “looked fine”
What This Book Covers
-
The Deterministic Sandwich: a pattern for wrapping AI
calls in
Prep -> Model -> Validation - Mission Objects: a way to turn vague requests and chat into typed, executable contracts
-
Validators: how to turn “looks fine” into deterministic
PASS/FAILchecks - Circuit breakers: techniques for keeping loops finite and auditable
- Governance at scale: practices for protecting the graders from the system they grade
About
Architects of Intent is an online book on Software Development as Code (SDaC). SDaC is the name I’m giving to building reliable, auditable GenAI systems with deterministic context, gates, and verifiable loops.
Written by Jóhann Haukur Gunnarsson, an Icelandic systems architect. I’ve built systems where audit trails and failure modes matter, including in finance, and I’ve learned the hard way that “looks fine” is not a strategy. This book is a practical synthesis of the loops, gates, and evidence that help make AI-assisted changes boring in production.
I offer consulting on the systems and practices in this book through
LumiLoop,
where we help organizations ground AI in persistent organizational
reality and build compounding operating loops.
I’m also a co-founder at
AI
Green Bytes, which is focused on sovereign,
sustainability-minded cloud infrastructure for AI inference. I expect
that as governed loops make AI reliable enough to run more real work,
inference demand will rise with them, and infrastructure quality will
matter just as much as model quality.
Connect on
LinkedIn.