You’re the CEO of a paper company.

A customer ordered a set of memos, stapled. Easy enough: print paper, staple the pages.

Another customer asks for your print paper for their product catalog. You think about what you know about the business. Yep, something glossy is great. Do that.

Another customer calls and wants you to consult on their trade show booth: user experience, collateral, layouts, handouts. That’s a whole lot of work. Time to call Sue.

Another customer says they want a hippo. You sell paper. Time to ask them what the heck they mean.

The paper business illustrates an important framework for how to think about building an AI agent:

1. Staple paper (Make a tool call) — a deterministic action with no judgment. Verify the results and don’t assume success just because there was no error.

2. Decide to use glossy paper (Reason inline) — when it’s a simple judgment that can be made with context on hand.

3. Plan the trade show booth (Call a sub-agent) — when specialized instruction and different tools are needed. Make the context minimal, and check the output.

4. Asked for a Hippo (add a fallback) – If things break at any point, add a fallback: ask the user.

Leave a comment