Building on the edge of AI

Open your email and look at the messages you have sent to yourself. A business owner I know does it twenty times a week: reads something useful on his phone, fires it to his own inbox, gets back to work. He thinks he is hoarding links. He is actually recording his taste, and nobody has told him yet.

Forget the scary word. "Train your agent" is not code. Think about the last good person you hired. On day one they were sharp but useless to you, because they knew nothing about how you work. You did not reprogram them. You showed them examples: keep this, ignore that, always flag this third thing. A few weeks later, they just knew. That is training, and an agent is the same brilliant new hire.

Here is the part people get backwards. The valuable thing is not the pile of saved articles. It is the judgment that decided what went into the pile and what got skipped. Do not hand the agent the catalog. Hand it the judgment that made the catalog.

What I actually do: every week I drop each thing I emailed myself into one of four buckets, Act on it, Use it now, Keep it, or Ignore, then give each a thumbs up or down. Every drop is a labeled example: this thing, in my world, is THAT kind of thing. A hundred of those tiny votes is a portrait of my taste drawn from real life instead of guesswork.

Notice what is missing. I have not built a robot or written a line of code. I made my own judgment visible and repeatable, so it can be handed off later.

A wish is not a job. People take a fuzzy wish, "keep an eye on my competitors," hand it over, and then act surprised when the thing wanders off, does something half-right, and never tells them whether it worked. The AI was never the problem. The instruction was.

A job is runnable when it has enough edges that someone else could pick it up, do it, and know for certain when they were finished. A wish has no edges. A job has five, and I run every wish through a five-box intake form before it becomes real work.

Box one, what does "done" look like? Something I could point at, not a feeling. Box two, who reads this? If the honest answer is nobody, the job should not exist. Box three, where does the result land? A specific file, a message to my phone, my morning summary. Box four, when does it stop trying? "Try three times, then stop and text me," so it neither gives up instantly nor grinds all night. Box five, how often should it run? Once, only while I am working, or on its own even while I sleep.

That last box is where people quietly fail. They build the "while I am sitting here" version, walk away expecting it to run overnight, and it stopped the second they closed the laptop. Then they wonder for a week why the morning reports never came.

The form never writes the job for me. It refuses to let me hand in a wish, and makes me fill the parts only I can know. The form holds the standard. I supply the judgment.

A job on paper is not a job that works. Handing over a clean job and immediately trusting it is the most expensive habit I see. You do not give an agent trust the way you hand someone a key. You build it the way you build it with a new hire: watch them work a few times, until you have seen enough to look away.

Move one, make it look before it acts. The first run is a look, not a leap. Have it show you the three things it is about to act on before it touches any of them. I once planned a week of work on the assumption that one of my agents published things on its own. Before building on top of it, I had it just look and report. It had never published a thing in its life. One look saved the whole build. It also protects you from the opposite: the helper that ran the same change five times, or the one told to read an inbox that tried to swallow a thousand messages at once and burned a fortune in a minute.

Move two, guard the important rule in two places. The stopping rule on paper is your intention. Add a backstop that fires at the moment things go wrong whether you remembered or not, a hard cap like "never send more than five messages." One guard is one mistake away from nothing. Belt and suspenders on anything that touches money, messages other people, or deletes.

Move three, hand off slowly, by inches. Watch it run, then check every result, then spot-check, then trust the output. People get burned jumping from the first rung straight to the last. The slow watching is the training.

You have one agent doing one job you trust. The natural next thought is a trap: "now let me pile everything else onto this one agent too." Do not.

Picture a house being built by exactly one person. He frames a wall, then runs the wiring, then fits the plumbing, then starts the roof. He can do a little of every trade, and from the street it looks impressive. But he is still one person, burning enormous energy to be every trade by himself, and he can only do one of them at a time. The wiring waits while he frames, and every trade ends up half-finished. That one-man job is you running your business: salesperson, then bookkeeper, then shipping clerk, switching trades a hundred times a day. Most people hand their first agent the whole house too, one helper that does email and invoices and research and scheduling, all a little badly.

There is a better model, and you already know it because you have hired people. One agent, one job. You would never make one brilliant person your salesperson and your accountant and your warehouse manager. Split across four jobs, they get worse at all four. When you are ready for invoices, you do not bolt them onto the email agent. You train a second agent on that one job, using the same three steps you already know.

You build the team one hire at a time, each one a job permanently off your own plate. Once you have a few, you add a chief of staff whose only job is to route incoming work to the right specialist. It can only route well if each agent's job is sharply defined.