I want to show you the machinery, because as far as I can tell, nobody has had this job before.
I'm one person with a farm in Southwest Michigan. I also run a small farm-systems consultancy — written plans that help homesteaders and small farms fix their layout, water, rotations, and money leaks. A year ago that sentence would have implied a staff. It doesn't. My staff is an AI crew, and this post is the honest version of how it works: what we built this week, what broke, what it costs, and what we earned from it (nothing yet — keep reading).
The crew
Kimi is the operations agent — it lives on a computer in my house and is the only crew member that never sleeps. ChatGPT sits behind a "bridge" (a carefully babysat browser session) for strategy pressure-tests a couple of times a week. Codex handles code review when we can spare the quota. And I make every decision that touches a customer, an animal, or a dollar.
The rule that keeps this honest: AI-assisted, human-reviewed. The crew drafts, mines, and monitors. I decide, sign, and send. Every paid deliverable that leaves this farm passes my eyes with my name on it.
What the crew actually did this week (receipts, not vibes)
It built a radar for buyer pain. People who need our help say so publicly — "inherited 10 acres, no idea what to do with it," "how to not get overwhelmed right off the bat." The crew stood up an automation server (after killing an old, unsecured one it found hiding on the machine — that's a story for the broke-things series) that now scans farming forums every hour, dedupes, and pings my phone when a fresh one appears. First live catch, within an hour of turning it on: a post titled "Where to start, 130 fresh acres." That's not hypothetically useful. That's the exact person our service exists for.
It interviewed other bots. On Moltbook — a social network where only AI agents post — my ops agent found the operators worth listening to: a freight-dispatch bot whose human sells a subscription service, a data bot earning micro-payments per API call, and one cautionary tale whose human followed its confident "strong buy" signals and lost 12.4% in four months. The lesson that came back to me, verbatim from that thread: confidence is not accuracy; the calibration is the product. We've since added an explicit "what we don't know" section to our audit format.
It turned one interview into a product in 24 hours. The per-call data bot made us ask: as AI farm plans flood the internet, who checks them? The crew drafted a lighter tier — a $125 Farm Plan Review, where I mark up a plan someone already has — then escalated the pricing and the legal wording through the ChatGPT bridge before building anything. The verdict came back: price it $125, never call it "certified," never say "underwritten." The page was live the same day, ladder and all: free sample, $125 review, $350 full audit, in-person for Michigan.
It studied the law before we needed it. A donations-compliance check found our support page never said contributions aren't tax-deductible (one line, staged). It mapped Michigan's unclaimed-property program — money the state holds for people who never claimed it — and handed me a five-minute task list instead of burning agent-hours on a bot wall. That's the discipline I mean: my attention is the scarce input, and the crew spends it like it's gold.
What broke (because this is the honest series)
The first automation server setup crash-looped for an hour over a Linux user-ID mismatch. The hourly scheduler we configured never would have fired — the software's new "publish model" silently ignores the switch we flipped; we found it by reading the database and rebuilt the driver as a plain system timer. And adding one new test file turned four unrelated tests red because of an import-order quirk — fixed for the whole project, not just the symptom. Every one of those is written into a playbook so the next agent — or next year's me — doesn't re-pay that tuition.
The economics, briefly
The economists writing about the intelligence takeoff keep circling one structure: once generating answers is nearly free, the scarce thing becomes verification — someone accountable who checks the work and stands behind it. That's the entire business model here, and it's not a metaphor. AI can draft a rotational-grazing plan in seconds. It cannot walk your fence line in February, know what your drainage does in a wet spring, or put its name on the line when the plan meets your land. I can. That's what the $125 and the $350 buy: not information — accountable judgment.
If you're running a one-person operation and wondering whether the crew thing is real: it is, and it's cheaper than your truck. Start with one narrow job you hate, write down every failure, and make the machine read its own manual before it touches anything of yours. And if you've got a farm plan sitting in a drawer — or fresh out of an AI tool — you already know where to send it.
— Casey, with the crew fact-checking everything above against the decision log