Day 3 of 100: My AI System Was Mostly Doing Work About Itself. So I Turned It Off.

On Day 1 I told you my AI system was finally draining work on its own. What I didn't realize yet was that a lot of the work it was draining was work about itself.

The machine was eating my time

I built this whole engine to take chunks of work and run them in the background. The idea was simple. Get the pipeline built first, and then everything else in the business would flow.

So I made system work the number one priority. And the system got really good at creating more work. A bunch of it was more system work. And a good chunk of my own time was going to it too.

Meanwhile the stuff that actually mattered that day wasn't always what ran. I run four companies. There's always something on fire, some project that got ignored a while back and now needs attention, and a few things I've been meaning to do. My priorities change every day. A background queue doesn't know what's on fire today.

So at the end of the day yesterday, we called an audible and turned it all off.

Experiment 2: one session, one result

This morning I tried something different. Instead of "go run all these chunks," I'd say exactly what result I wanted and how it's supposed to work. Then one AI session would re-research everything from scratch, with zero assumptions: what's there, what's not, what was done, what wasn't, where the gaps are. Then it builds a plan and drives it, start to finish, while I'm in the conversation.

It felt way more efficient. I honestly can't tell yet how much of that is the method and how much is just that I can see what's happening instead of it running in the background. Jury's still out.

The problem was I had three of those heavy sessions going at once, and by mid-morning they'd burned through one AI account's entire usage window together.

Experiment 3: one session to run all the others

So by lunch I tried one more thing. One "conductor" session that controls all the other sessions, spreading the work across multiple AI accounts so I don't burn through my usage limits so fast.

It kind of worked. The heavy work (the stuff that touches several projects or links a few pieces of software together) ran the rest of the day without running any one account dry. That part was great.

But it was clunky. Questions would pop up in new windows instead of the conversation I was already in, and none of those windows would open on my phone. I was stuck at the desk all day playing whack-a-mole. Even the conductor's own summary at the end of the day said it mostly built itself. Kind of embarrassing.

Where I've landed (for now)

So here is where I've landed for now. One session per result, and I stay in it. That session hands the heavy stuff off to our other AI accounts in the background so nothing runs dry.

I keep a separate window open for each one, each its own Claude Code session, and I can pull any of them up on my phone. Honestly that helps me keep my head straight with four companies going.

I also like to work from my phone so I can actually walk around the neighborhood while I answer questions or whatever. It lets me get some exercise in so I'm not stuck at the computer all day. I try to get my 10,000 steps in, or at least hit my 500 calorie a day goal. I think it's important to not sit all day. So if the questions can't get to my phone, it doesn't really work for me.

The goal hasn't changed. Ideally I wouldn't be in sessions at all. I want the AI to just do the work, and do it properly. But there are decisions I have to make, forks in the road where I have to choose my own adventure, so for now I'm still needed in there.

If you want to follow along, follow me on X. If you run a company and you're thinking "I want this in my business," tell me about it here.

-Matt