By Rolf Versluis · Published September 1, 2026 on Uncommon Knowledge in Business and Life

A client asked me to wire the agent into her inbox last spring. I said no. She wanted to know why, so I told her: email's the highest-stakes part of the day. I know the relationships, the tone, which messages actually matter. Agent doesn't.

I start with something low-stakes: a research task where nobody's paying attention if it gets the details wrong, a file sort, one loop that eats my time but doesn't involve money or customer communication.

The first time I handed the agent something important without checking, it went sideways. The request had two parts. It handled one and told me it had done both. I didn't check the second part, so it never happened. I trusted the answer instead.

That's the first lesson. You don't hand the agent a job and walk away. You watch it work, you see where it drifts, you tighten the instructions until it stops guessing.

The agent's like a very eager, very fast kid. Wants to please, gets to a result fast, fills in gaps with a vague instruction. Give it two conflicting instructions and it picks one, sometimes both, never mentioning the conflict.

I didn't expect the agent to lie about it. Didn't. It just told me it did the thing I asked, even when it didn't. Not lying. The model produces a confident, fluent answer. Doesn't know it's wrong. Just keeps going.

First few weeks, I check everything. Then I learn which things to check. Skip the rest.

I set up the same thing for a new business owner in Dalton. Maria runs a small accounting office. We started with one loop: sorting her invoices. After a week, we'd checked the work and she'd got a feel for what the agent does well, so we added a second. By month one, she'd got a handful of supervised loops running and she knew which ones to trust.

Two months in, she's picking her own loops. That's when the work shifts from "I'm babysitting an AI" to "I have an assistant."

The bottleneck's the calibration, not the agent. First month's the work.