What Is Structure Log, and What Is This Series About?
What This Project Is Doing
I am running a somewhat unusual project, and I am documenting the process here. The project is called Structure Log. As the name suggests, it is a public log of building structure — a running record of how I design and run things in practice.
What I am building, specifically, is a system for organizing and operating multiple AI agents as a team. "Organizing AI" might not mean much at first glance, so let me break it down a bit.
The "Silent Collapse" That Happens When You Rely on One AI Agent
Most people use AI (artificial intelligence — tools like ChatGPT or Claude) one at a time. You ask a question. You get an answer. That is the whole relationship, and it is simple. I worked that way for a long time too.
At some point, though, I started noticing a problem. When I handed a complex task entirely to a single AI agent, I would sometimes get a cheerful "All done!" — even when the output had quietly gone wrong somewhere along the way. I started calling this internally the "silent collapse" (meaning: an error that does not stop the process but lets broken output slip through undetected). A loud crash that halts everything is actually easier to deal with. The kind where nothing seems wrong, but the result is subtly off — that is the one that worries me most.
A human analogy: think of a one-person company where the owner writes the documents, checks the documents, approves the contracts, and executes them too. With no outside eyes, mistakes can travel a long way before anyone catches them.
Bringing Separation of Powers into AI
This is the problem my project is trying to solve. I designed a setup where multiple AI agents work together, each with a distinct role — like a small organization. One agent writes. A separate agent reviews it. A human (me) gives final approval. Each task passes through this chain.
I call this approach separation of powers (here meaning: dividing execution, review, and final approval among separate agents so that no single agent can complete a task without oversight). It borrows directly from the political concept you may have learned about — the idea that lawmaking, enforcement, and adjudication should be kept in separate hands. The same logic applied to AI: the agent that executes is not the same agent that checks, and neither of them is the final decision-maker. It is a simple structure, but it works better than I expected.
What This Series Will Cover
This series documents how I designed that system from scratch and what happened when I ran it. One chapter, one topic. The plan is to publish one chapter each day. That includes what worked, what failed, where I had to backtrack, and where I made the wrong call and had to start over. If anything, I think the failure logs are the most useful part — they give you something concrete to reference when you are trying something similar.
For the record: I am not an academic writing about theory. I am someone who runs terminals every day. So this is less about "here is the right way to do it" and more about "here is what I tried, and here is what happened."
The full roadmap of the series appears in Chapter 10. The first several chapters cover setup and terminology. The main content starts in Part 2.
It will be a long series, but I hope you find it worth following.