The $50 AI Company

A stack of cheap AI agents can feel like running a small company for about the price of a few streaming subscriptions. It is not a company until the work is written down.

I learned that the hard way: at the peak I ran about 12 agents and really used one or two, a story I tell in I Built My Own Agent Graveyard. What I kept is one agent that handles many tasks with skills. That only works if each skill has a proper page behind it.

The $50 in the title is a public frame for "cheap enough to try", not a bill. The point is how the work is organised, not what it costs.

Standard work, not personas

Anyone who has spent time on a sort floor knows standard work. It is the written, visible way a job is done today: the steps, the order, the checks, the time it should take. It is not a straitjacket. The people doing the work improve it. But on the Monday before Christmas, when half the shift is new, it is the reason a first-timer can do the job safely and well.

A good agent role is standard work. A persona prompt is not. "You are a brilliant, proactive chief of staff" tells a model how to sound. It does not tell it what to do, what to never do, or when it is finished.

The test I borrowed from the floor is the new-casual test. Could someone on their first peak shift follow this page? For agents the question becomes: could a different model pick up this role tomorrow and do the job? If the answer is no, the role lives in a conversation, not on a page, and it will drift.

Five lines per role

Every role in my setup, which now means every skill my one agent can call, has the same five lines. Here is the real one for my morning brief, lightly trimmed:

• Trigger: 6am daily, or when I ask for it.
• Inputs: calendar sync, open loops, the training plan, a few flags from my own finances. Nothing from work systems.
• Output: one brief I can act on in ten minutes, written to a file and then emailed.
• Must not: invent meetings, copy private journal material into anything shareable, or send anything on my behalf.
• Done: I have decisions and load in front of me, not a summary of my inbox.

Five lines is not much. It is enough to make the role portable, reviewable and testable.

The must-not line is the one that earns its place. In a morning brief, one made-up meeting does more damage than any number of good summaries. A nicer prompt will not hold that line. The words "must not invent meetings" on the page, and a test that fails if a meeting appears that is not in the source, will. When a role goes wrong now, I fix the page first, then the tooling, then the model settings. Usually the page was the problem.

Three keys stay with me

Most things can be delegated to an agent. Three cannot: send, spend and publish.

Send means anything that leaves my world with my name on it: email, messages, posts. Spend means money, including the quiet kind like a new paid tool. Publish means anything public, including this site.

In my setup those keys are enforced by missing capability, not by instructions. The agents can draft email but have no way to send it. They can suggest a purchase but cannot make one. They can prepare an edition but cannot deploy the site. When something goes out, it is because I clicked.

That is not distrust. It is where accountability sits. A company where nobody owns the send is not a company.

The stop rule

A company also needs a way to stop doing things. Agent setups rarely have one. Roles get added because they are cheap, and they stay because nobody notices them.

My rule is simple. Every role gets a review date. At the review, I ask one question: did I use what it produced in the last two weeks? If not, it is retired. Not rebranded, not given a better prompt, retired. Its runbook goes into an archive folder in case I want it back.

The stop rule is what keeps a cheap company cheap. Without it, the cost is not the subscription. It is your attention, spent reading output nobody needed.

What this means at work

Organisations are about to do the same thing at scale. A copilot for every team, an agent for every process, each with a friendly name. Some of that will be excellent. Some of it will be twenty group chats.

The questions to ask are the same ones a good operations leader asks of any process. Where is the standard work written down? Who improves it? Who holds send, spend and publish? What is the stop rule? If a vendor answers with mascot names and a demo, ask again.

If you want a bigger team than your headcount, do not collect characters. Write the roles.

Which AI role do you rely on most, and could a different model pick up its page tomorrow and do the job?