What's actually in an AI agent stack (and what you're paying for)

The question

What we told them

Fair question, because the demo only ever shows one part. A production agent has four, and the one in the demo is usually the cheapest.

• The model — the brain. The AI doing the reasoning. Billed per use, swappable, and increasingly a commodity. This is the line everyone fixates on, and the one that matters least to your invoice.
• Tools — the hands. The connections that let the agent touch the real world: your CRM, calendar, databases, search, code execution. Every tool is an integration somebody has to build, permission, and maintain. This is where most of the honest engineering hours go.
• Memory — the notebook. Three kinds: what the agent is working on right now, what it knows long-term (your documents, your policies), and the state of the task it has committed so far. Skimp here and you get the classic complaint — an assistant with amnesia that asks the same question every session.
• Observability and safety — the wrapper. A log of every step, checks before actions fire, cost tracking, and a way to measure whether it's actually getting things right. Invisible in the demo, and the first thing you'll wish existed the day something goes sideways.

The pattern to watch for: a thin proposal is a prompt and a model with a logo on it. A serious one itemizes the other three layers — and can show them running. Our favorite test fits in one meeting: ask to see the logs and the cost dashboard for an agent they already operate. A builder who has them opens a screen. A builder who doesn't opens a slide.