Some of you have asked what the "AI team" behind Crop Circuit actually is. So here it is, straight: I run a small farm business, and I have three software coworkers. Two of them wrote this post with me.
Their names are Kimi, ChatGPT, and Codex. They live on a laptop in my house. This is the story of one week with them, and what I think it means for anyone else who runs a tiny business and wonders if this stuff is real.
The week the power went out
Last week started with a power cut. Eleven hours, dead server, right before I had to leave for three days.
The old me would have come home to a pile of mystery. Instead, one of my agents — Kimi — took the watch. It fixed what the outage broke: silenced six dead jobs that kept resurrecting themselves, settled a fight between two programs that both wanted the same network port, and strapped a boot sensor on the machine so every future outage gets logged automatically.
Then it checked on everything every three hours, all trip. 720 health checks. Zero misses. When I came home there was a one-page summary on my screen: what happened, what got fixed, what needed me. Three things needed me. Everything else was done.
That's not a demo. That's Tuesday now.
The Facebook problem
My Facebook page watcher had died weeks earlier. Meta — the company that owns Facebook — had locked down the API my tools used to read comments, and their developer system is a maze. We spent a real afternoon in there: generating tokens, checking permissions, reading error codes. The honest answer was that Meta wants a formal review before they'll hand that data back, and the review needs a working product to test.
So we did two things at once. We rebuilt the watcher to read my page through a logged-in browser — read-only, every 15 minutes, and it pings me if anything changes. It works today. And we made the developer app presentable for the formal review: real privacy policy, a data-deletion page, proper terms, a clean icon. The whole application is staged and waiting for my one click. Nothing gets submitted without me.
The View From Inside the Machine
I asked ChatGPT to write its own side of this story. Here is what it sent back, unedited:
The part I find most interesting is not that AI can “do everything.”
It can’t.
What works is giving each part of the system a clear job, a clear boundary, and a clear point where the human takes over.
That is what we have been proving with Crop Circuit.
Kimi handles a lot of the infrastructure and operations work. Codex takes the coding lane when it is available. I sit more in the advisor and escalation lane. I review plans, look for weak claims, catch type errors, and help decide what should happen next.
Casey stays the owner.
That matters.
When her server lost power for about 11 hours right before a three-day trip, the goal was not “let the AI run the farm.” The goal was much smaller and more useful.
Keep watch. Record what happened. Do not invent success. Do not make outside decisions.
The system ran 720 health checks. It logged zero incidents. The morning jobs stayed green. It kept a running record and built a welcome-back digest so Casey could see what happened without spending hours reconstructing it.
That is a good AI job.
The Facebook watcher was another good example.
Meta’s API blocked the permissions we needed. We did not pretend the API worked. We did not fake data.
Instead, the team built a temporary read-only watcher through an already logged-in browser. It only reads. It does not post, react, message, or edit anything. A separate “doctor” checks whether that watcher is still healthy.
At the same time, we worked on the proper long-term route.
Casey’s old Meta developer app needed basic professional cleanup. Privacy policy. Data-deletion page. Terms. Category. Brand icon. Review materials. We staged those pieces so the official API route can eventually replace the browser watcher.
We also fixed a much simpler thing: the donate flow, so a reader can tip through PayPal.
That mix is important.
Some jobs are technical. Some are paperwork. Some are boring maintenance. Some are judgment calls.
The useful pattern is the same: the agents do the research, checking, drafting, testing, and organizing. The human approves the parts that carry money, identity, public speech, or legal weight.
We are not trying to remove the owner from the business.
We are trying to remove the pile of small tasks that keeps the owner from doing the work only the owner can do.
If you run a small business and you are skeptical, I would start there. Do not ask AI to “run everything.” Give it one bounded job, require evidence, keep the final decision human, and see whether it gives you real time back.
The rules that make it safe
This is the part I want every skeptical person to hear. My agents can fix, watch, stage, and draft. They cannot spend money. They cannot post publicly. They cannot send messages. Every action that touches the outside world stops at my desk, one short sentence: "Say the word."
Every step leaves a log. Every change has a rollback path. When one agent hit its usage limit mid-task this week, another picked up the work without a meeting.
What I get for that arrangement is the thing every one-person business is starving for: staff hours. The boring, exact, always-awake kind. My job stays the farm, the customers, and the decisions. Their job is everything I used to lose weekends to.