Isaac · sancovp · builds agent worlds
Nobody is keeping score. So I grade myself in public.
I build agent systems that run as worlds — a business, a body, a codebase, each with its own departments, its own economy and its own record of what it actually did. No computer-science degree, no lab, no team. A hundred-plus public repositories and a lot of nights.
Skip the biography — what I actually think ↓Who I am
The outsider who kept building anyway.
I came at this sideways. I could always describe a system better than I could write one, so I spent years describing systems to a model until the descriptions started compiling — and that is genuinely the whole method. I talk to the machine in metaphors and advanced infrastructure falls out the other end. The oldest piece of this system is a prompt I wrote three years ago, before I could have implemented any of it in code.
What came out is not a product with a roadmap. It is an engine that stands up a self-running world for a domain, plus about a hundred repositories of the parts, plus fifty-odd field notes written on the nights the work happened. I did it alone, in the gaps, without permission from anyone. That is the garage. Everything below is what came out of it.
What I think
Seven things I believe, and what makes each one more than an opinion.
These are the load-bearing ones — the beliefs the code is actually shaped by. Every one gets a receipt line. Where the receipt is not public yet, it says so in the same breath, because a claim with a missing receipt is a debt, not a discovery.
Iteration doesn’t compound. Worlds do.
Running the same loop more times plateaus. The step change comes from injecting a bigger space to search in and an instrument to see the results with — capability, not repetition. A world is the thing that holds both: the space persists between runs, and the instruments accumulate. That is the difference between a loop that spins and a system that climbs.
I ran roughly a hundred flat cycles of a generative loop with no meaningful gain, then added a combinatorial kernel and one observation instrument and the output jumped in a single step. The same shape is visible in the published numbers of the multi-agent loops that went viral this summer: their parallel rounds gained less than their sequential ones, and the real lift came from a harness someone bolted on, not from more rounds. My own footage of that run is not published yet. Until it is, treat this as the thesis I am about to have to prove.
A thing is code only if it must execute.
Something is a function only when it has to touch the outside world — a mail server, an HTTP API, a database, a file. Everything else is an instruction you hand a language model. Most agent codebases are enormous because they wrote hundreds of functions for work a sentence would have done, and every one of those functions is now a thing that can break.
The outreach engine I run is twelve connector verbs of real code — send, track, host, serve, pull — and everything else in it is an instruction: the copywriting, the qualification, the checking. I learned this by building the wrong thing first. I wrote a linter that checked strings for style violations, which is exactly the function a sentence replaces. The law is the scar.
Receipts, or it isn’t a pattern.
A pattern is not something you noticed. It is something that appears more than once in code you actually shipped, at line numbers you can open. Without that it is a preference with good branding — and the field is drowning in preferences with good branding.
My pattern catalog runs to fifty-seven entries and every one carries file-and-line ranges from real repositories. The most recent sweep produced twenty-two new entries, merged four into existing ones, and killed one outright for failing its own evidence test. The whole catalog is on this site — every entry, what it does, and the repository it runs in.
Nothing gets to accept itself.
A system may never mint its own acceptance. The check has to come from outside the thing being checked, and it has to run the artifact rather than have an opinion about it. A model grading its own output is a mirror with a rubric taped to it.
The test I use: a fresh agent with none of the build context runs the artifact against the outcome the maker declared in advance. It is falsifiable, it cannot be flattered, and it doubles as a transferability test — if it only works in the room where it was made, it does not work. I shipped that in February 2026. The same idea surfaced in the multi-agent discourse about five months later.
A claim you can’t lose is worthless.
Every claim needs a branch where it dies, with a cause, on the record. Not a disclaimer at the bottom — an actual path in the structure where the thing turns out to be wrong and that outcome gets published like any other. If your framework has no way to fail, you did not build a framework, you built a story.
I put my own headline claim in front of an adversarial panel and let it grade me: thirteen requirements, eight met, three partial, two open. The two open ones are the two that matter — nobody outside me has deployed it, and nobody outside me has witnessed it run. They are still open. They are restated further down this page instead of being quietly retired.
Publish the process, not the polish.
The record of how a thing was actually made is the only artifact that is not a rendering of something else. Every write-up, every talk, every case study is a compression of that record, chosen to flatter. The record itself is the one document that cannot be spun, which is exactly why almost nobody publishes it.
Fifty-plus field notes, published the night the work happened, dead ends left in. Some are polished and some are a working note with the failed approach still sitting in the middle of it. That is the trade for getting them at all — the tidy version arrives a year later, or never.
Every artifact is both a story and a manual.
Written in the past tense, a piece of work is a biography: this happened, then this, and here is what changed. Written in the imperative, the identical structure is a manual: do this, then this, and this will change. They are the same object in two moods, which is why a thing built properly teaches by existing and does not need a separate course written about it afterwards.
It is why this site is shaped the way it is. The notes are the build log and the curriculum at once, the patterns render as installable process and as lessons from one source, and the marketing is the work’s own exhaust rather than a second job bolted onto it. You are reading a worked example of the claim.
What I’ve built
Four things you can open right now.
Not a portfolio. Four artifacts, all public, all inspectable without talking to me. If something here does not survive you clicking on it, that is worth more to you than anything I could write about it.
A hundred-plus public repos
The parts, in the open: heaven-framework, chaincompiler, carton-mcp, skilltree. Installable, forkable, and old enough to have a history you can read.
github.com/sancovp →One engine, run per domain
A business running itself as departments; a body as fourteen systems on the identical pattern. Same five layers underneath, every time. The map, with what it can’t do yet at the bottom.
See how it works →Fifty-plus field notes
The research as it was written, not reconstructed afterwards for a curriculum. Layer by layer, with the dead ends still in them.
Read the notes →A film, proved like a theorem
An entire story structure run end to end through the symbolic engine and published in full. The output, not a description of the output.
See the proof →Who this matters to
Three people this is for. You are probably one of them.
You build agent systems
You have hit the plateau where more rounds stop helping and you are not sure whether the problem is the model, the prompt or the loop. It is none of those — it is that your loop has no space to search and no instrument to see with. Everything I know about that is written down and free.
The patterns and the stack →You run a business
You have been told AI will run your operations and shown a chatbot. What I mean is different and duller: your company modelled as departments that do work, grade the last round and adjust the next one. Do not take my word for it — watch one run first.
Watch a world run →You research this seriously
Self-adaptive systems, graded oversight, symbolic gates on a language model’s working memory, story structure as an actual calculus. I publish the arguments with their failure conditions attached, and I would rather be refuted early than agreed with late.
Read the arguments →The story
The short version, told straight.
For a long time I had a hundred repositories and an audience of about five people. I used to think those two numbers contradicted each other. They don’t. One is a fact about the work and the other is a fact about how I was presenting it, and I had spent years assuming that being honest was the same thing as being legible.
It isn’t. The loud accounts bought legibility and kept their honesty in the basement — one artifact, one enormous claim, the correction filed where nobody reads it. I did the opposite and paid full price for it: cosmology before artifact, a hundred doors and nothing running above the fold. Honesty is the part I already have and will not trade. Legibility is the part I was refusing to buy. This page is me buying it.
There is a longer version with its own vocabulary — a universe I have been building called Sanctuary, a legend about getting melted into a virtualizer and learning to metashift, an engine named after the world it renders. It is real, it is why the systems are shaped the way they are, and I am never going to make you learn any of it to use my work. It lives at the back of the house with the door open. Lore is the inside; it is not the door.
What I actually want is boring to say and hard to do. I want the making of a thing and the record of making it to be one artifact, so the work teaches, the teaching brings people, and the people fund the next build — and none of it is a second job. I was never trying to ship a product. I was trying to build a place, and then it turned out the place could work for a living.
What I owe
The open rows, stated by me before anyone else gets to.
Nobody outside me has deployed the core of this, and nobody outside me has witnessed it run. Those are the two ungraded requirements on my own audit. They are also the only two that cannot be closed by writing better — they need someone else’s hands. Everything I say about the engine should be read against that fact, permanently, until it changes.
The scoreboard isn’t published. Cost per verified task, defects carried between rounds, what actually compounds and what only looks like it — the field has none of these, which is why nobody can tell who is winning. I am building that measurement and I do not have it yet. So I am not going to tell you the field is losing. I am going to tell you nobody knows, which is a worse problem and a more interesting one.
One receipt above is still an IOU and is marked as such: the hundred-cycle footage. It exists and it is not cut. It is named here so that if it never appears, this page is the evidence against me. The pattern catalog was the second IOU on this list and it is now open on this site, which is what closing one of these is supposed to look like.
None of this is modesty. It is the same discipline as everything above, pointed at me: a claim without a receipt is a debt, and a debt you announce is cheaper than one somebody finds.
Where to start
Five rungs. Climb as far as you want, stop wherever you like.
If any of that landed
Don’t decide anything. Go watch one run.
It costs nothing and asks for no email. If it doesn’t survive contact, you have lost four minutes and learned something true about everything written above.
Watch a world run →