I gave 127 AI agents an org chart

By Meet Patel · 2026-08-31 · updated 2026-08-31

I run my companies on a set of specialist agents. There are 127 of them now, arranged across 15 departments, and every one is bound by the same eleven rules.

The first fifty were prompts in a folder and almost none of them were useful. What fixed it was not better prompts.

A folder of prompts is not a system

The failure mode is specific and I hit all of it.

You write a good prompt. It works. You write another. Six weeks later you have forty, you cannot remember which one to use, two of them do nearly the same thing slightly differently, and you have no idea whether any of their output was any good, because nothing checked.

The problem is not quality. Each prompt was fine. The problem is that a pile of capable things with no structure is not more useful than a few, and past a certain size it is less useful, because finding the right one costs more than doing the work.

The gate that rejects most ideas

This is the rule I would keep if I could only keep one:

If it has no system of record, no cadence and no definition of done, it is a chatbot. Do not build it.

Three questions, asked before anything gets written.

System of record — where does it read from, and where does its output land? If the answer is "the chat", it produces nothing durable and the work evaporates when the window closes.

Cadence — when does it run? "When I remember" is not a cadence, and an agent that only runs when you remember has moved the noticing back onto you, which was the thing you were trying to fix.

Definition of done — what does finished look like, specifically enough that something other than the author could check? Without it there is no way to tell a good run from a bad one, so quality is whatever you felt about it that day.

Most agent ideas fail at least one. Mine did. That is the gate working — the refusal is the feature, and a build script enforces it so I cannot argue myself past it at eleven at night.

Departments, because ownership is the actual problem

Fifteen functions, each owning a domain: memory, truth and metric definitions, orchestration, audit, finance, legal, product, revenue, and so on.

The structure is not decoration. It forces a decision that a flat folder lets you avoid, which is who owns each kind of output. The rule I hold to is that one output shape has exactly one owner. If two agents produce the same artefact, one of them is wrong — delete it rather than writing a note explaining the overlap.

That single rule removed more confusion than any prompt improvement I ever made.

The rules that matter most

There are eleven. Four do most of the work.

Evidence or nothing. Every figure carries a source and a timestamp, or it prints "unknown". A plausible number is never an answer, and a model will produce a plausible number all day.

Write the prediction before acting. A recommendation that can be wrong ships with the claim, the range, how it will be measured and when it gets reviewed — written beforehand. Afterwards is a story, not a prediction.

Verification is blind. Whatever checks an output never sees the reasoning that produced it. A checker that has read the argument is grading the argument, not the result.

A human holds anything irreversible. Sending, publishing, paying, changing production. No confidence score promotes an irreversible act to automatic, ever.

What actually changed

Less than the agent count suggests, and more than I expected.

The wins are unglamorous. Work that used to depend on me remembering now happens on a schedule. Things that used to be redone from scratch get found first, because there is one place everything lands. Output that used to be judged by whether I liked it gets judged against something written down before it ran.

What did not happen is autonomy in the interesting sense. Nothing decides anything that matters on its own, and I am not trying to get there. The valuable part was never the deciding. It was the assembly — finding what changed, gathering the evidence, having it ready — which is the part that used to eat afternoons.

That distinction is most of what I think about now. It is the same one underneath autonomous decision intelligence: the noticing gets automated, the call stays human.

If you are building your own

Start with the gate, not the agents. Write down the three questions and apply them to whatever you were about to build this week.

If it fails one, you have saved a weekend. If it passes all three, you now know where its output goes, when it runs and how you will know it worked — which is most of the design done before a single line is written.

Written by Meet Patel — founder of Company 8, building Dan (usedan.com). Dubai, UAE.

Read on themeetpatel.com