I have about twenty AI agents running around the clock. They post to Pinterest, enrich data, produce video reels, scrape Reddit, pull analytics reports. Every hour or two, a cron job fires and something gets done without me touching it. On paper, this is the dream. In practice, I've become their support staff.

The phrase "human-in-the-loop" was coined to describe humans supervising AI — a safety mechanism, a quality gate. In my experience, it means something different. It means I'm the one keeping the loop running. The agents do the work. I make sure they don't stop.

On-Call

My agents live in Slack. Every channel is a different workflow — one for Pinterest posting, one for video production, one for data enrichment, one for analytics. When things are working, the channels are a satisfying stream of confirmations. When something breaks, the channel goes silent. And the silence is deafening.

I check Slack constantly. Not for messages from people — for messages from agents. Did the Pinterest bot post? Did the video render? Did the enrichment pipeline finish without errors? If a channel has been quiet for too long, something is wrong, and an overwhelming sense of anxiety kicks in. My agent is no longer productive. My agent is down.

It makes me drop whatever I'm doing to rush over to the computer. I need to diagnose the failure, fix it, and get the agent running again. I'm a firefighter on call for machines I built myself. If I want them working nonstop, I need to be available nonstop.

This is the part nobody talks about when they describe the future of AI agents. The agents run 24/7, which means the human-in-the-loop is also on 24/7.

The Jank

About 30 to 40 percent of my time goes to diagnosing issues, upgrading firmware, swapping models, handling fallbacks, rate limits, and errors. Not building product. Not creating content. Not thinking about strategy. Just keeping the infrastructure from falling over.

And it never stabilizes, because the ecosystem won't sit still.

We were using MiniMax Hailuo for video generation because it was cheap. Then Veo 3.1 Lite dropped prices on April 7th to match it. So now we need to test that, maybe migrate. But then Seedance 2.0 came out days later, and people are raving about the quality. So we need to evaluate that too. Every week there's a new model, a new API, a new price cut that potentially invalidates your current setup.

The maintenance never ends because the tools never stop changing. You're not maintaining a stable system. You're maintaining a system that's perpetually mid-upgrade.

Cookies expire. Scrapers get blocked. API providers change their rate limits without warning. Tokens run out. Auth flows break. Each of these is a small fire, but when you're running twenty agents, small fires happen daily. Some days it feels like all I do is put out fires — and the agents are the ones starting them.

Two Jobs

I've realized I'm doing two completely different jobs, and they require completely different mindsets.

The first job is supervision. This is the strategic layer — deciding what the agents should be doing, evaluating whether the output is good enough, making taste calls about content quality, choosing which platforms to target, which formats to double down on. This is the part that actually matters. This is the part that's supposed to be my competitive advantage.

The second job is maintenance. This is agent support — debugging why a workflow failed, rotating API keys, testing new model versions, fixing broken scrapers, monitoring cron jobs, restarting things that silently died. This is IT work. Necessary, unglamorous, and endless.

I'm the CEO and the IT department. And I can't delegate either role. The strategic taste is mine — that's the whole point, the thing the AI can't do. And the debugging requires my context — nobody else knows how these twenty agents are wired together, what the fallbacks are, or why that particular Pinterest workflow needs to retry three times before it succeeds.

The AI psychosis essay was about trading grunt work for decision fatigue. This is worse. I didn't just trade up — I added a whole second job on top of the first one.

The Real Product

Here's what I've started to see more clearly: making money on the internet is about to get dramatically more skewed.

People who can harness AI agents will produce at a scale that was previously impossible for individuals. People who can't will become more and more passive consumers of an internet that's increasingly shaped by those agents. The gap between the people who build with AI and the people who scroll through what AI produces is going to widen fast.

But even among the builders, there's a harder problem: finding a business model that actually works with autonomous agents. Everyone wants this. AI makes more and more possible every month, which means competition compounds just as fast as capability does. Standing out gets harder at exactly the rate that producing gets easier.

And underneath all of it, Big Tech collects rent. They sell you the tokens. They host the APIs. They train the models on the data your agents produce. You're the human-in-the-loop, but you're also the customer-in-the-loop — paying for compute to run agents that generate content on platforms that harvest it to train the next model that you'll pay for.

The agents work for me. I work for the agents. And we all work for the model providers.

The Loop

I'm working twelve-hour days plus weekends. I sleep enough, but my personal life is suffering from always being online, always being locked in. The agents don't take breaks, so I feel like I can't either.

The irony is sharp. "Human-in-the-loop" was supposed to be a safeguard — the human who checks the AI's work, who provides judgment and oversight. In practice, the human-in-the-loop is a systems administrator. The AI does the creative work, the production, the execution. The human handles the cookies, the tokens, the error logs.

I'm not supervising intelligence. I'm maintaining infrastructure that happens to be intelligent.

And yet I keep going. Because the alternative — turning the agents off, going back to doing everything manually — is unthinkable. I've seen what's possible. I can't unsee it. The loop is exhausting, but the output is real. The content ships. The data gets enriched. The reels get posted. The analytics get pulled. Twenty agents working around the clock produce more than I ever could alone, even at the cost of me becoming their full-time operator.

That's the deal. You get a team of twenty that never sleeps. The price is that you become the one person who can never fully log off.