Free — no email, no strings
The patterns, written plainly.
13 short pieces on how these systems actually work, and the long field-note feed behind them. Take what you need.
Start reading →Or jump straight to a concept
Frameworks
A way to think about the problem. These hand you the shape, not the build.
Compounds Or Rots
Any system that feeds its own output back into itself is on one of two trajectories. The fork is a single property, and most systems default to the wrong one.
Read →The Altitude Ladder
There's a ladder, it shows up twice, and most people are lower on it than they think. The most honest move on it is usually "don't climb."
Read →The Four Slots
Anything that has to navigate complexity and not fall apart is the same four things plus the connections between them. The slot you can't answer cleanly is the bug.
Read →Build In Verified Layers
Complexity is built like a tower, and you don't advance a layer until the current one gives you an observable signal it can hold weight. "Felt ready" is not a signal.
Read →Drift Is The Default
Everything that generates, drifts. Correctness is not something you install once — it's something you maintain by returning, and the interval is the quality.
Read →Right Is Not Fluent
Two completely different guarantees you can ask of an AI system, routinely conflated. More context makes a fluent system more convincingly fluent; it does not make it right.
Read →Forge Your Own Vocabulary
You can't discuss what you can't name. Your mind reaches across domains before it speaks — and the shared structure between those reaches is the name you were missing.
Read →Case studies
One real thing that happened, what it cost, and what was actually learned.
The $5,000 Product That Was Ten Minutes of Work
I Paid an AI Agency $5,000. Then I Found Out What They Actually Did. Strip the branding off a typical "AI agent" and the setup is about ten minutes of work.
Read →The codebase that got safer to edit every time it was touched
A repo isn't a pile of files — it's overlapping invisible surfaces. Here's why AI-written codebases rot, and what it looks like when one stops.
Read →The validator that won't let the AI be wrong about Susan
Susan processes invoices every Tuesday. Your AI knows none of this — and will happily generate a confident, plausible workflow that has never existed.
Read →The architecture that drifted for months, diagnosed in two minutes
An architecture drifted for months and nobody could even name the problem. Everest expeditions and wingsuit jumps turned out to be the way in.
Read →The compiler that turns an ordinary thing into structure
My dad always has snacks. The shallow read is "he's thoughtful." Chase the real answer and you find a clean structure under the costume — and that's repeatable.
Read →The agent that stopped drowning when it only saw the next turn
Give an agent three hundred tools and it gets worse, not better. The flip is embarrassing once you see it — and it changes the failure mode, not just the failure rate.
Read →Field notes
The long feed — 51 notes published as the work happened, filed by concept.
The whole feed
Every field note on agent engineering, filed under the seven levels — prompts, tools, context, harnesses, admissibility, concentration, emergence — and the concepts that cut across them.
Browse all 64 →Flat vs Tree
A bug hiding in every AI “skills” folder — and the paper about it. They named it “progressive disclosure,” then shipped it unable to disclose.
Read →Who writes these
Seven things he believes, each with the receipt attached — and, at the bottom, the debts he names before anyone else gets to.
Meet the maker →Receipts, not testimonials
Don't take my word for it — take the code.
No logo wall, no invented case studies. Four things you can open right now.
The engine, on GitHub
heaven-framework, chaincompiler, carton-mcp, skilltree — MIT, full commit history.
github.com/sancovp →A world running in your browser
The engine drawing its own explainer. Your scroll is the clock.
Watch it →What one run produced
A live local-business site, built end to end by the engine. Not a mockup.
Open it →The field notes
Published as they were written — including the ones that record a mistake.
Read them →Where this sits