Early in building my AI systems, I handed a task to an AI agent and walked away.

Bad idea.

When I came back, the task was done. But the bill was $416. For a simple automation workflow that should have cost maybe $20.

What went wrong?

The agent had no memory of the previous session. So it started from scratch — re-discovering everything it had already figured out the last time. Every re-discovery costs money. Every credential failure costs money. Every loop costs money.

I wasn’t watching. So nobody stopped it.

Here’s what I learned:

  1. AI agents have no memory between sessions unless you give them one. At the end of every session, write a handoff document. What was built. What’s working. What’s not. The exact next step. One page. Five minutes. Saves hundreds of dollars.
  2. Set a spending cap before every session. No exceptions. It sounds obvious. It isn’t, until you get the bill.
  3. Be in the room. AI agents are powerful. They’re also literal. They will do exactly what you said — not what you meant. If you’re not watching, you find out after.
  4. Simple builds don’t need an AI agent. Agents are for complex, multi-step autonomous work. For simple tasks, a human (me) with the right tools is faster, cheaper, and less likely to loop for two hours on a credential error.

$416 is cheap for a lesson this valuable.

I run every AI build session differently now. And I haven’t had a runaway bill since.

— Warren