The Model Doesn't Do Anything — and Most Agent Explainers Skip That

Morgan Linton published a walkthrough of how AI agents work aimed at semi-technical professionals, and one section is worth the entire click.

The piece states plainly that when a model "uses a tool," it doesn't actually call anything. It emits a block of JSON — structured text describing what it wants to do — and the runtime handles execution. This is the single most widespread misconception about agents right now. People imagine the LLM reaching out across the internet, pulling data, running functions. It isn't. It's autocompleting a tool-call schema. The runtime does the work. Linton doesn't bury this in a footnote or hedge it — he puts it front and centre with a code example, and it lands.

That's where the piece stops pulling its weight.

The orchestration loop section presents clean pseudocode — while not done: call model, execute tools, continue — and treats it as the explanation. But the reason agent engineering is hard has almost nothing to do with that loop. It's everything the pseudocode leaves out: what happens when a tool returns an error. How you cap iteration depth. How you prevent the model from re-calling a tool that sends an email or deletes a record. How you handle the model hallucinating a tool name that doesn't exist. These aren't edge cases you encounter later — they're the first things that break when you move past a demo.

There's also a quiet bait-and-switch in the framing. What's described is a tool-calling loop: user asks a question, model selects tools, results come back, done. That's a useful pattern, but it's not what most people mean when they say "agent." Planning, goal decomposition, self-evaluation, backtracking — none of it appears. The piece delivers "how tool-calling works" but bills itself as "how AI agents work." Those aren't the same thing, and conflating them is how professionals end up overestimating what off-the-shelf agent frameworks will do for them.

The tool-call mental model is worth passing along. Just know that the hard part starts exactly where the article ends.