My Mac was running 155 scheduled jobs and I could only explain 54 of them

Last week the fan on my laptop had been running hot for days and everything felt like it was moving through mud, so I finally sat down and took inventory of every background job on the machine.

The thing that runs work on a schedule on a Mac is called launchd. It's what wakes a script up every minute, or every night at two in the morning, and runs it without asking you. I've been building agents for months, and every time one of them needed to run on a clock, I handed it a launchd job. The count came back at 155.

I had started all 155 of them myself. 101 were pointed at scripts or project folders that don't exist anymore, so they'd been firing into nothing for weeks and nobody noticed, including me.

Six of them were worse than useless. Three jobs from my approvals system were exiting with code 78, which means their config was broken, so they died on startup and launchd restarted them, and they died again. Every 60 to 120 seconds. For nine days. My content engine job was exiting 127, which means it was calling a command that isn't installed on this machine anymore. A leftover Postgres job was failing on every single attempt. That was the fan: thousands of Node processes starting, crashing, and getting resurrected, until the memory compressor had nothing left to give.

None of that is the interesting part. The interesting part is that I built every one of those jobs on purpose, and I couldn't tell you what most of them were for.

There's a well-known study of London cab drivers who memorized 25,000 streets to pass the Knowledge. Their hippocampus, the part of the brain that builds spatial maps, measurably grew. After they retired and stopped driving, it shrank back. A 2020 study out of McGill went the other direction and looked at ordinary drivers instead, and in a small follow-up group, the people who leaned hardest on GPS lost the most ground on their own sense of direction across three years.

The one I keep coming back to is older. In 2011 a team at Columbia ran a set of experiments on what happens when you know a fact is going to stay saved on a computer. People remembered where the fact lived instead of what it said. They kept the folder and dropped the contents.

My machine had become the folder. 155 jobs, all mine, all filed somewhere I'd stopped looking.

I want to be careful here, because there's an easy version of this essay that ends with "do it yourself," and I don't believe that. Handing work off is the entire reason to run agents. I'm not going back to writing my own cron entries by hand, and the 54 jobs that survived the audit are doing real work I'd be worse off without. Delegation isn't the failure.

What I'd quietly stopped doing was the deciding.

At some point over the last few months, my prompts got shorter and shorter until a lot of them were basically "fix it." I'd paste an error, type fix it, read the first two lines of what came back, and move on. That works often enough to be dangerous. Claude is good, so the output usually looks right, and looking right is enough to keep you from checking. Every one of those 101 dead jobs was created in a moment where I knew exactly what I wanted, and then I never went back, because going back would have meant holding the whole picture in my head again, and the machine was already holding it for me.

Except it wasn't. Nothing was. The jobs were running and no one, human or otherwise, knew why.

So there are two things now. The first is a kill list. 101 jobs are getting removed, and the six failing ones get fixed or deleted, no third option. That's an afternoon of work and it's the easy half.

The second is a rule I'm holding myself to before I hand anything else to an agent. I have to be able to say out loud, in a sentence, what I want the job to do and why it needs to exist. Not type it. Say it. If I can't finish the sentence, the job doesn't get made, because a job I can't describe is a job I won't remember, and in six weeks it'll be another line in an inventory I have to run with the fan screaming.

The prompt is where the thinking is supposed to happen. That was always the part that stayed mine.