The Three Things I Got Wrong Running My Company on AI Agents

90 days ago I handed most of my company’s daily work to AI agents.

Three things I got wrong. None were the ones I braced for.

1. I thought the hard part was the tools.

It wasn’t. I wired up plenty. The wins never came from a better model. They came from writing down the calls my team makes on instinct, the ones nobody had put into words. An agent doesn’t need more tools. It needs your judgment, made explicit.

2. I gave it access before I gave it boundaries.

I optimized for speed, so the agents could reach almost everything. That is how you get a confident action nobody asked for, aimed at something that matters: a customer, production, real money. The fix was boring. Make the agent fail closed. When it isn’t sure, it stops and asks. Slower some days. Safer every day.

3. I assumed knowledge would travel.

It didn’t. Each agent was sharp at its own task and blind to what the last one just learned. Context doesn’t carry itself. The real project was never smarter agents. It was the memory layer underneath, so what one learns, the next inherits.

That last one is the real lesson. I stopped building better agents and started managing them. An agent is a new hire who never sleeps and never learned your judgment. Most of the work is management, not magic.