For a few weeks I had one agent and one rule: ask it to do the thing, and the thing gets done. This is the version of AI work that fits on a billboard. It is also the version that quietly falls apart the moment two things are happening at once.

The agent could scaffold a project, write the code, fix the tests, update the docs, and tell me it was finished, all in one breath. Impressive, until I tried to answer an ordinary question a week later: who decided this project was a lab and not a tool, where exactly did the build stop and the review begin, and which step was responsible for the bit that later broke. One agent had done all of it, which meant no part of it was anyone’s job in particular.

One agent doing everything is nobody doing anything

The problem was not capability. The single agent was good. The problem was that “do everything” is not a role, it is the absence of one.

When the same worker scaffolds, builds, documents, and signs off, there is no seam to inspect. Every output arrives already blended together, and the reasoning is a smooth paste with no edges. I could not point at the moment a decision was made because nothing in the system was shaped to own a decision. Accountability needs boundaries, and a generalist has none by design.

A worker with no boundary cannot be held to one. That is not a personality flaw of the agent. It is the shape of the job I gave it.

So I had speed and no structure, which is a familiar trade until you notice it is not a trade at all. You can have both. You just have to stop hiring one genius and start hiring a department.

I tried to fix it with better instructions

My first move was the obvious one. If the single agent blurred its jobs together, I would write a longer brief telling it to keep them separate. Scaffold first. Then build. Then document. Then review yourself. Please.

It worked in the way that asking yourself to be more organised works. For about a session. The agent held the structure while the instruction was fresh and dissolved it the moment a task got interesting, because the structure lived in a paragraph it could choose to reinterpret, not in the shape of the work itself. I was asking one role to remember to behave like six.

What got weird: the roles started arguing, and that was the point

The thing that changed my mind happened by accident. I had split the work across a few separate, narrowly briefed agents, and one of them, the reviewer, refused to sign off on work another one had produced. The builder had done something reasonable; the reviewer said it failed the taxonomy and safety checks and would not pass it.

A single agent would never have done this. A single agent agrees with itself, always, because there is no second party in the room. The disagreement was only possible because two roles had genuinely different jobs and different definitions of done. The friction was not a bug. It was the first sign that the work had stopped being one opinion and started being a process.

What worked: a small department of bounded roles

So I stopped designing a smarter agent and started designing an org chart for one person to run. The unit of design became the role, and a role earns its place only if it can state one job and one stop condition. If it cannot say when it is done, it is not a role, it is a mood.

Today the working roster is six. Each operates on the factory itself, and each hands off to the next:

  1. Scaffolding. Turns a rough idea into a classified, documented, validated project container. Stop condition: the container exists and passes its checks. It does not write product code.
  2. Build. Implements scoped work inside a container that already exists. Stop condition: the scoped change is done and validated. It does not decide what the project is.
  3. Documentation. Keeps the technical trail honest: decisions, session logs, traceability. Stop condition: the change is recorded where the next session will find it.
  4. Walkthrough. Maintains the plain-English explanation for a human who was not in the room. Stop condition: a non-technical reader could follow what happened.
  5. Review. Checks code, docs, validation, taxonomy, and safety, and is allowed to say no. Stop condition: a verdict, with located findings.
  6. Promotion. Assesses whether a lab should graduate, stay, or be archived. Stop condition: a recommendation, never an automatic action.

There is a seventh seat I have left empty on purpose: an architecture-map role, planned but not yet hired, because the work has not demanded it often enough to justify the boundary. The roster grows by need, not by ambition. A role you add before you need it is just a costume.

The handoff order matters as much as the roles. Scaffolding hands to Build, Build hands to Documentation, Documentation to Walkthrough, then Review, then Promotion. Each handoff is a place where one role’s output becomes the next role’s input, with a written trail in between. The seams I could not find inside the single agent are now the most useful part of the system, because every seam is a place I can stand and ask what happened.

The reusable part: name the role, then name where it stops

If you take one thing from this, do not take my six roles. Take the test that produced them. For any piece of agentic work, ask two questions: what is this role’s single job, and how does it know it is finished. If the second question has no clean answer, you have not found a role yet, you have found a wish.

The roster is the product here, more than any individual model behind it. Swap the model underneath and the department still works, because the department is made of boundaries, jobs, and stop conditions, and those do not live in the model. The test travels even when the roster does not.

The fantasy was that AI would do the work. What actually happened is more boring and more durable: I stopped being a person with a very fast assistant and became a person running a very small, very disciplined department, where the most valuable employees are the ones who know exactly when to stop.

What Next

Building the Factory While Standing Inside It

The roster needed somewhere to live. So I pointed the scaffolding role at the factory itself, and over one weekend it built its own children.