Farm Agentic Operations

I'm standing in a field with no power, no computer, and no internet, and I've just remembered the newsletter didn't go out. Authoring it on my phone would be insane. So I don't. I open a session I left running at home, on my full machine, and the newsletter goes out from the field anyway.

Running a farm is running a complex machine. We had multiple sales channels, a weekly newsletter, a CSA roster, records of who showed up and who didn't, and a pick list that depends on what I picked last week. On top of that came a relentless stream of one-offs. All of it happened at home, at night, in front of a computer. Small businesses usually deal with this by settling for a mediocre offering, because you have to cut the things a larger firm could do. We weren't willing to make those trade-offs, and the refusal was killing us. It took all of our time and it hampered our growth.

Early AI tools had one pitch for this: give us your documents and we'll answer questions about them. Useful, but not what I wanted. Then, working with Claude Code early on, it clicked. This wasn't a coding tool, and it wasn't a chatbot. It was an agent that sits in a folder on your machine and interacts with the files around it. It can read them, write them, search them, reorganize them, and learn from them. So the move was to build a folder an agent could live in, and then put these new agentic powers to work running the farm. I set up a repository with everything an agent should know about my farm: the website, the Shopify store, the grants we'd applied for, the research we'd done, the apps we'd built to support our operations, and the APIs we relied on.

Now when I want to write a newsletter, I have a conversation with my Claude agent. I hand it the list of what we've decided goes out that week and it drafts an email in our voice, from MJML templates that live in the repository. It keeps a history of every newsletter we've ever sent, so it knows when it's about to repeat itself. It pulls in the menu and recipe ideas, checks that every link that should be there is there, and puts in the right slugs without being told. Coming up with the slugs alone used to take almost twenty minutes. Now it's an afterthought. Every week we need to know who's coming to pick up their CSA share, so we get a pickup list built from live order state in Shopify, run through a custom system that merges in the swaps when someone moves from one week to another and the notes about who's missing. It renders to HTML and prints to a physical copy we use at the farm. How many shares do we pack? Answered. Prices need to change partway through the season? Done. Need to get our cash sales and market booth fees into QuickBooks? Ask. Every ad campaign, every crop plan, every pick list comes through this folder. I won't say it removed the work, but it has easily cut the time these tasks take by 90 percent. We nudge it the last 10: a little more of this, a little less of that, then give it the go-ahead.

If that sounds easy, well, it was and it wasn't. The great thing about working with these agents is that you just ask them to do things, and sometimes it simply works. Other times it doesn't, and getting to results that were both acceptable and consistent took real iteration. At first we were handing it ad-hoc instructions, which meant the same job got done three different ways. It loved to invent claims: that we're certified organic, that our vegetables are picked hours before every delivery, that we sell products we don't even carry. For a farm whose customers shake our hands at pickup, those aren't typos. They're claims we can't stand behind. The system had to become trustworthy.

The biggest part of the answer turned out to be teaching the model the way you'd teach a new employee, in three steps. First, SOPs, which in this case means Claude skills. We started by writing down what the farm is and what it offers. We gave it sources of truth and skills around them that it would always read into its context before writing anything a customer would see, then copywriting guardrails, then a how-to for every system the agent touches. As good results started coming out, we set them aside as examples of what success looked like, and that became a self-reinforcing flywheel: better examples, better content, better examples. Second, authority, granted in stages. At first it created drafts that I looked over heavily and sent manually, whether that was an email, a new website page, or whatever else. Then it got to where it would make a draft and, for the most part, I'd approve it and push a button. As the confidence kept building, we let it send. Third, verification, so I could confirm it had done what it thought it did. This is the step a lot of people miss. They think they can engineer a perfect, infallible system. The reality is that it's almost impossible to build things that can't break, and when they do break, you need to know so you can fix it. So we added checks and push notifications for anything that isn't working. That caught things like the time the agent set up an email correctly but the scheduled job that was supposed to send it hit a provider outage, and the newsletter never went out. With the checks in place, any time there's an outage we know about it and can route around it.

There's a principle in that last step that I think generalizes well past the farm. You can spend enormous engineering effort trying to build a system that never breaks, or you can build checks that tell you when something is out of whack. Agents are stochastic. You can never be certain they haven't done something you didn't expect. So detection is worth more than robustness, even when robustness is what you're aiming for.

The skills have kept getting better, and somewhere along the way the folder stopped being a repo and became an operating system for our farm. We've now run a full season on it. The farm operates like a much larger company, with a responsiveness and a quality of content and service that two-person businesses don't get to have. We spend our time in the field, where the work can't be automated, instead of in the office, where it can. I'm still standing in that field. My phone buzzes. The newsletter went out. I get back to figuring out how to grow a better beet.