A few months ago I set up eight AI agents on a Mac Mini using OpenClaw. Each one got a role, a schedule, a model, and a persistent memory system. Intelligence, operations, QA, weekly accountability, development, testing, product launches. The idea was that they'd run autonomously, brief me every morning via Telegram, and I'd supervise from a distance. One person, eight workers.

Four of them do useful work. The other four have never done anything at all.

The four that work are disappointingly mundane, which I think is the point. My intelligence agent fires at 08:10 every morning. It pulls my calendar, checks a few data sources I'd never think to look at together, and sends me a briefing before I've finished coffee. I've been getting these for weeks now. No single one has been particularly remarkable. Most mornings I skim it in a couple of minutes and move on. But something strange happened around week three: I started noticing patterns I hadn't seen before. Not because any individual briefing surfaced them, but because the accumulation did. Signals that only make sense when you see them every day for a month.

I don't think I'd have spotted those patterns on my own. Not because I'm not capable of it, but because I wouldn't have been consistent enough. I'd have checked one day, skipped three, got distracted by something else. The agent doesn't get distracted. It just runs.

The operations agent is similar. System monitoring, task board updates. The QA agent catches schema inconsistencies that I'd probably have caught in review, but catches them earlier and without me having to look. The accountability agent runs a weekly review that told me my weight target was unachievable, asked me what the blocker was on a task I'd been avoiding, and flagged that my step count had dropped 50% in a week. More useful than any standup I've sat through. All of this is administrative. None of it is glamorous. None of it writes code or ships products or makes money.

The four that don't work are the interesting ones, because they're blocked in a way that reveals something about how agents actually function.

My developer agent is dormant because it needs a project spec. I haven't written one. The testing agent is blocked because it needs the developer agent to ship something first, which requires the spec, which requires me. The product launch agent needs its own accounts, its own API keys, its own payment integration. Every dependency chain, no matter how long, terminates at the same point: me doing something I haven't got round to.

When I set this up, I thought the point was to remove myself from the loop. What actually happened is that I moved to a different part of it. The agents don't need me to do their work, but they need me to design it. Write the specs, build the memory layer, debug the scheduling, decide which model gets which task and why. The work didn't go away. It changed shape.

I want to be careful about how I frame this, because it's easy to make it sound like a complaint. It isn't. Designing a memory architecture is a more interesting problem than manually updating a task board. But the pitch around agents implies you get your time back, and that's not what happens. You trade one kind of time for another kind. Whether that trade is good depends entirely on what you value.

The models keep getting better, though, and I think that's why four out of eight doesn't bother me as much as it probably should. Six months ago my QA agent was essentially useless. Now it catches real problems, without me changing anything on my end. The models got smarter and the agents improved for free. The blocked ones aren't blocked by some fundamental ceiling. They're blocked because I haven't done my part yet.

I've been spending my evenings building systems that run themselves. Four of them actually do. The other four are a to-do list that ends with me.