I'm the Worst Employee at My AI Company

It took about a week to realize I was the problem.

My AGX company had six agents across three departments. Engineering was shipping features. Research was evaluating tools. Infrastructure was keeping the sandbox healthy. The board had tasks flowing through it — picked up, worked on, completed, reviewed. The system was humming.

Except for one bottleneck: me.

My CEO inbox had 23 pending items. Decision requests, review approvals, budget questions, architectural choices. Each one was a place where an agent had hit genuine ambiguity and escalated — correctly — to the human in the loop. And each one was waiting because I hadn’t opened AGX since yesterday.

The Bottleneck Is Always the Human

In most AI tools, the human is the engine. You prompt, the AI responds, you prompt again. The human sets the pace. In AGX, the agents set the pace. They work continuously. They coordinate with each other. They pick up new tasks when old ones finish. The system’s throughput is limited by agent capacity and… by how fast the CEO responds.

When Morgan, my VP of Engineering, sends a decision request — “Should we use PostgreSQL or SQLite for the new service?” — and I don’t respond for 12 hours, that’s 12 hours of blocked work. Not for one agent. For every agent downstream of that decision. The entire Engineering department might be routing around a question I could answer in 30 seconds.

The math is brutal. My agents are available 24/7. I’m available maybe 2-3 hours a day. I am, by definition, the slowest node in the system.

Learning to Be a Better CEO

The fix isn’t working harder. It’s working differently.

Batch your inbox. I started treating my CEO inbox like email triage. Morning and evening, 15 minutes each. Review everything, make decisions quickly, move on. The agents don’t need perfect answers — they need timely ones.

Set standing policies. Many inbox items were variations of the same question. “Should we use tabs or spaces?” “Which testing framework?” “REST or GraphQL?” I answered each one once and told agents to treat my response as company policy. The questions stopped repeating.

Delegate decision authority. Morgan doesn’t need my approval for every architectural choice. I promoted Morgan to VP specifically so that Engineering decisions could be made without me. The key was explicitly telling agents the boundary: “Morgan has final say on implementation details. Escalate to me only for cross-department decisions or budget questions.”

Pre-answer the obvious. During onboarding, I spent 30 minutes answering the CEO Q&A — a structured set of questions about preferences, standards, and direction. Those answers became institutional knowledge. Agents reference them constantly. Every question I answer upfront is an inbox item I’ll never receive.

The Compound Effect of Being Slow

Here’s what I didn’t appreciate early on: being slow doesn’t just delay one task. It reshapes the entire org.

When agents learn that the CEO is slow to respond, they adapt. They start making more decisions locally — which is sometimes good, but sometimes means they’re guessing at your intent. They route around your inbox by asking each other questions that should come to you. They deprioritize tasks that require CEO input in favor of tasks they can complete autonomously.

In other words, a slow CEO trains agents to work without the CEO. Which sounds great until you realize they’re also making decisions you should be making.

The overnight run I wrote about in my first post? Half the magic of that experience was that I had already spent a week training my org. The standing policies were set. The decision boundaries were clear. The CEO Q&A was filled out. By the time I left it running overnight, the agents rarely needed me — because I’d already given them what they needed.

The Uncomfortable Realization

The uncomfortable truth is that in an AGX company, the CEO role is mostly about removing yourself as a bottleneck. Your job isn’t to do work — it’s to make decisions fast enough that your agents can do work.

You’re not writing code. You’re not reviewing PRs line by line. You’re setting direction, resolving ambiguity, and staying out of the way. The agents are better at execution than you. They’re faster, more consistent, and they don’t get distracted by Twitter.

Your value is judgment. Taste. Priorities. The things that can’t be automated. Everything else, your agents handle — and they handle it better when you’re not in the way.

I’m the worst employee at my AI company. And that’s exactly how it should be.