In Lesson 1 I described an agent as software that works through a task in several steps and uses tools along the way. That definition is fine as far as it goes. Today’s lesson was about what actually happens inside those steps, and it turned out to be simpler, and a bit stranger, than I expected.
I had pictured an agent as something that makes a careful plan and then carries it out. It doesn’t work like that at all.
It’s a loop, not a plan
An agent doesn’t map out the whole task in advance the way a script would. It runs a cycle: think, act, look at the result, think again. One step at a time, reacting to whatever just happened. The pattern has a name you’ll hear in AI circles: ReAct, short for “reason + act.”
This is closer to how most of us cook without a recipe. You taste, add a little salt, taste again, decide it needs lemon. Nobody writes the whole sequence down first. You let each result tell you what to do next. That’s flexible, and it’s also how you end up with a soup that drifted somewhere you didn’t intend.
The model never touches the tools
This was the part that surprised me. The AI model itself can’t check a calendar or send an email. When it wants to use a tool, it writes out a structured request, something like “call the calendar tool with these details.” The software wrapped around the model makes the actual call, then pastes the result back into the conversation. The model reads that result and decides its next move.
So the model is a bit like someone directing a job over the phone. It can’t press any buttons. It can only say which button to press and then listen to what happened.
That arrangement explains both sides of agents. It’s why they can do things instead of just saying things. It’s also why they go sideways: every lap of the loop is a fresh chance to misunderstand. A small misreading on step two gets carried into step three, and by step six the agent can be confidently working on the wrong problem. Anyone who has played the telephone game, or followed a group chat that took one message the wrong way, knows how that goes.
Where MCP fits
For an agent to use a tool, the two have to be connected somehow. Before MCP, every tool needed custom wiring for every agent, which was a mess. MCP (the Model Context Protocol) is an agreed-upon plug shape: any tool that speaks it works with any agent that speaks it.
The easiest comparison is USB. Before it, every printer, mouse and camera came with its own odd cable. Now you plug things in and mostly don’t think about it. MCP is trying to be that for AI tools.
The one-line version
An agent is a model trapped in a loop with a toolbox, deciding one step at a time.
I like that sentence because it takes some of the mystery away without making agents sound trivial. “Deciding one step at a time” is also a fair description of most of my own days. The difference is that I usually notice when a step went wrong. The question worth asking of any agent is who, or what, does that noticing for it.
Try this today
Pick something small you do in several steps, like planning dinner from what’s in the fridge or finding a time to meet with a friend. Notice how often you look at a result and change your next move. Then imagine handing that task to software that can only ask for tools by writing requests and reading back what it gets. Where would it be most likely to misunderstand? That spot is where you’d want to check its work.
Next lesson: why the answer costs more than the question.