DoNotEnergize · Episode 01

A Model Is Not an Agent

An answer is not an access badge. A model and the system that lets it act are different things.

The film

2:45 · AI-assisted educational film

Three ideas to keep

A model generates outputs from the input available to it. The surrounding application determines which tools it can use and how actions are carried out.

An agent system can use model outputs, tools, and feedback to work through steps toward a goal. Some steps can be fixed in code; others can be chosen by the model.

Permissions belong in the software and services that execute actions. Asking the model to be careful does not itself enforce an access boundary.

Ask before you connect a tool

What can it read?
What can it change?
Which actions require a person’s approval?
How will we check the result?

These are companion notes, not a verbatim transcript.

Further reading: Anthropic: Building effective agents (opens in a new tab).

Next: What Is a Token?

Full spoken transcript

The words of the published lesson. The key ideas above remain companion notes.

If it can answer your question, that does not mean it has permission to roam the building. An answer is not an access badge. Start with the model. It takes input and produces output using patterns learned during training. Ask it to draft a plan and it gives you a response. Now the agent. People use that word differently. Here we mean a larger software system that uses a model to work toward a goal. It can choose a step, use a tool, check the result and decide what to do next. The model helps guide those decisions. The surrounding software runs the tools. Think of the model as a brilliant consultant in a locked room. Building an agent is like adding a work order, a toolbox and an access badge. Which seems like an excellent time to read the change ticket. But remember, that is an analogy. The model did not grow ambitions or a secret career plan. People built the system. They chose the tools and granted the access. They also need to decide when it asks for approval and when it stops. At work, a model might draft a plan for a software update. An agent might also check the system, request approval, run the update and inspect the result. That depends on which tools and permissions the surrounding system provides. So when somebody says, are we buying an AI agent? ask: Can it only answer or can it act? Which tools can it use? What can it access? How many steps can it take? Where does a human approve? And what stops it? A model may be brilliant. The system around it determines the blast radius. How far a mistake can reach. Next signal, what is a token? And why are we spending them like arcade coins? Until then, understand the machine, question the story and please read the change ticket before energizing.