← All thoughts

Agent-native operations

Running a seven-figure store with agents in the loop

I have been running Phone Loops for years. It sells phone straps and accessories online, it does seven figures, and for a long time it ran the way most small companies run: on a few people, a lot of tabs, and my memory.

Today most of the daily work is done by agents. Not one assistant answering questions in a chat window. A set of small machines, each with one job, reading the same state and reporting to the same place.

This is what I learned getting there.

Adding AI is not the same as being agent-native

The first thing everyone tries is to bolt a chatbot onto the business as it is. Ask it to write a product description, summarize the inbox, draft a reply. It helps a little. Nothing really changes, because the business still lives in people's heads and the agent has to be told everything, every time.

Agent-native means the reverse. You redesign the operation so that an agent can read where things stand and act on it without a briefing. The work is not in the prompt. The work is in making the business legible.

Rule one: if it only lives in a conversation, it didn't happen

Every task, decision and idea gets written to a file the moment it shows up. Stock levels, open purchase orders, product launches, what was decided last Tuesday and why. Structured, versioned, boring.

That sounds like bureaucracy. It is the opposite. Once the state is written down, an agent can pick up where another left off, a check can compare what should be true with what is true, and I can ask "where are we on this?" and get an answer built from facts instead of from someone's recollection.

Rule two: one agent, one lane

What runs at Phone Loops today, roughly:

Each of these has a written job description, a schedule, its own state, and a clear definition of what "done" looks like. None of them is clever on its own. The value is that they run every day whether or not I remember them.

Rule three: the agent comes to where I already am

I don't open dashboards. I live in Slack, like most operators. So every machine reports there, in a thread, in plain language. When something needs me, it shows up as a short list of options with one marked as recommended. I answer with a number.

That single format change did more for the business than any model upgrade. A decision that used to wait for me to sit down and think now takes ten seconds from my phone.

Rule four: measure what came out, not what ran

An agent that runs every morning and produces nothing looks exactly like a healthy agent on a quiet day. The only honest check is on the output: did orders ship, did the catalog change, did the email actually leave. I come back to this in Rigs, not scripts, because it is the most common way these systems quietly die.

What it changed

Revenue had been flat for a while. The point was never to fire people or chase a headline. The point was to stop trading my time for operations, so that the hours go to the next thing instead of to the same Tuesday checklist.

My calendar is mostly empty, and it is supposed to stay that way. The store runs. When it needs me, it asks, and the question is already half answered.

If you want to start

Don't start with a platform. Pick one loop that someone in your company runs by hand every week, write down what it reads, what it decides and what it produces, and build a small machine that does exactly that. Then make it report somewhere you already look. The second one goes faster. By the fifth, you are redesigning the business around them.


Running on too many manual hours?

Rigs, not scripts →