The AI agent itch

In the first five months of 2026, nearly 88,000 people in the US lost their jobs because of AI. That's already more than all of 2025. In May alone, AI was behind 40% of all layoffs in the country.

‍

Headlines like this are everywhere now. And when you read them — and then scroll past fifteen posts about AI agents that automate everything, replace teams, build empires — it creates this burning itch. This feeling that everyone around you knows something you don't. That you're falling behind. That you need to drop everything and start building agents right now, everywhere you can.

‍

I felt it too. Hard.

‍

I'm an engineer by education. I've been around software my whole career. I'm a science fiction fan who now gets to touch the things my favorite authors wrote about decades ago. And twenty years of running an agency taught me one thing above all else: look for practical use, not shiny use. I already work with AI constantly — ChatGPT, Claude, Perplexity, Gemini. Research, brainstorming, learning. I even have a small agent that gathers and enriches information on potential clients.

‍

But it felt like it wasn't enough. The itch kept saying: you need more. You need real agents. Autonomous ones. The kind that do things on their own without being asked.

‍

What I actually found

‍

About a month ago, I finished a public speaking course. Afterward, a small group of us went to a cozy little restaurant to celebrate. And while everyone was sharing laughs and favorite moments, I ended up in a two-hour conversation with one person — a young woman, fresh out of university with a degree in agricultural science of all things, who now builds (believe it or not)… AI agents for a living.

‍

For two hours we compared notes. Her practical experience building them. My twenty years of trying to make operations actually work. What struck me: almost every real agent use case I'd come across was about marketing. Content generation. Image creation. Publishing. Video. The internet is full of these. The few operational agents I'd seen were basically chat assistants doing predefined tasks when someone asked. Or they handled things like onboarding checklists and status summaries — which, as Claude later confirmed when I ran a proper brainstorm, matched my experience exactly — are closer to simple automation than a real agent's job.

‍

What I wanted was different. A truly autonomous agent. Something that could seriously lighten the operational load without me typing a command every time. Something that just ran. So I went deeper. I ran a research session inside AI itself — asked both ChatGPT and Claude to brainstorm with me.

‍

ChatGPT was agreeable. It proposed agent ideas that sounded impressive but would mostly add pure bureaucracy to a small team. More checks, more steps, more complexity dressed up as intelligence. Claude pushed back. And somewhere in that conversation, it said something that stopped me: "For a team of five to fifteen people, the only place an autonomous agent gives real value is when it notices what you'd miss — because you're not looking every day."

‍

Not reminding you. That's annoying. Not evaluating decisions. That's bureaucracy. Scanning the system and raising a signal that would otherwise be missed entirely. I sat there staring at the screen. Because I knew exactly what that was.

‍

We already had this

‍

Years ago, in my agency, we had a role we simply called "the controller." A person whose entire job was to notice what the rest of us would miss. Task statuses. Whether SOPs and Docs had owners. Whether Friday reports had actually been sent, or annual review dates were about to expire, or internal projects had stalled without anyone noticing. She'd walk around — sometimes literally — and surface what had gone quiet.

‍

This wasn't for client delivery. We had project managers for that. The controller watched everything else — sales, HR, accounting, marketing, internal projects. The stuff that slowly dies when nobody's looking. And here I was, in 2026, about to build an AI agent for something we'd figured out with a person and a checklist years before AI agents existed.

‍

The brainstorm with Claude ended with one more line that stuck: the main reason ops systems die isn't bad design. It's that people stop maintaining them. And that’s exactly it. That’s exactly the reason why most of my attempts to systemize our agency failed. Failed until I truly got that idea.

‍

SYSTEMS HAVE TO BE MAINTAINED.

‍

SOPs go stale. Roles stop being updated. Friday reviews get quietly dropped, escalation boards forgotten. The system doesn't collapse — it just slowly goes dark.

‍

An agent doesn't fix that. It amplifies whatever's already there. If there's a solid process underneath, an agent makes it sharper. If there's nothing underneath, the agent just adds another layer of complexity that someone — usually the founder — will have to babysit. My itch didn't need a new tool. It needed me to remember what I already knew.

‍

AI is genuinely useful. A well-trained model with deep context about your clients, services, team, and processes can cut communication overhead dramatically. That's real. But the breakthrough isn't the number of agents you deploy. It's how well you actually use the knowledge and experience your business already has. The rest — the endless feed telling us we need agents NOW or we'll be left behind — that's not strategy. That's FOMO with a subscription fee.

‍

Something to try this weekend:

‍

Pick the one process you've been itching to hand over to an AI agent. Don't ask Claude or ChatGPT "how do I build this." Ask the opposite: "Be brutal, not agreeable. What breaks if I automate this? What will I have to maintain forever? Is the process even solid enough to hand off?" The honest answer usually tells you whether an agent saves you time — or quietly steals it.

‍

So here's what I've been thinking about since that brainstorm: when you imagine an agent solving your biggest operational headache — is the problem actually about the work itself? Or is it about the fact that no one's really watching it?

That's the blind spot. I walked right into it with twenty years of experience behind me. The itch felt like it was about technology when it was really about attention. Hit reply and tell me what you find. I read every response.

‍

P.S. That controller we had? She wasn't technical. She wasn't a manager. She just paid attention to the things everyone else was too busy to notice. Turned out that was the most valuable role in the building.