Agents

Autonomy and authority aren’t the same thing. Agents can do the work. Restarting the system should still take my signature.

Match the model to the stakes. A small coding model is fine for light work and the wrong call for anything critical.

When a model swings from “this is nothing” to “build everything around this” in one reply, call it out and get a second opinion.

If an agent answers with a list of things for me to do, it has the roles backwards. I’m not the worker here. It is.

Separate the agents that decide policy from the agents that carry it out. Church and state.

Not every decision needs a frontier model. Let a cheap model judge whether the request needs an expensive one.

If your agent’s only memory is the chat window, it doesn’t have a memory.

I don’t want an AI that does everything. I want one that tells me exactly what it did.

A small model that knows its job beats a big model guessing at yours.

Run a model on your own machine once and you’ll understand what you’ve been trading for convenience.

The AI demo ends when the model says it worked. The product starts there.

Prompt engineering is a job title for a problem that should be solved in code.

Everyone’s asking what ChatGPT knows. The better question is what it will do when nobody’s checking.