An AI agent is a language model (the AI behind a chatbot) given three things: a goal, a set of tools it can use, and a loop. With those, it takes a real action, looks at what happened, and decides the next step on its own, instead of just replying once.
A goal
"Book a meeting with Sarah next week"
+plus
Tools
how it acts on the world
read_calendar()draft_invite()
+plus
A loop
act → look → decide the next step, repeat
🎯 Goal, not a script🔧 Tools = how it acts (calendar, email, search)🔁 A loop: it keeps going until done🛑 Stops for your OK on big moves
The plain analogy: a chatbot is a knowledgeable friend on the phone. They can tell you anything, but they can't lift a finger for you. An agent is a capable assistant in the room: give them a goal and they'll open the calendar, find a slot, and draft the invite. Same brain; the difference is the hands.
02 / 05 · The spectrum
"Isn't that just a fancier chatbot?"
The most common mix-up, and the one worth killing first. People call three different things "AI," and an agent is only the last one. Here's the spectrum.
⚠ The myth
"An AI agent is just a chatbot with extra steps." Or worse: "a sci-fi robot that thinks for itself and goes rogue."
✓ The reality
An agent isn't smarter than a chatbot. It's the same kind of model. The only difference: it's been handed tools and a loop, so it can act and then react to what it finds. It doesn't "think for itself" with a will of its own; it works toward the goal you gave it, with the tools you allowed, and (when set up well) it stops to ask before anything irreversible.
💬 Chatbot
Talks. Takes no action.
✓ Answers from what it knows
– Can't touch your calendar, email, files
– Replies once, then waits for you
› Great for: questions, drafts, explaining
⚙️ Automation
Same fixed steps, every time.
✓ Fast, cheap, totally predictable
– No judgment, one rigid path
– Breaks on messy or unexpected input
› Great for: when the path never changes
🤖 Agent
Given a goal, decides the steps.
✓ Picks which tools to use, in what order
✓ Adapts to what it finds along the way
✓ Handles messy, unstructured input
› Great for: judgment + several tools
The key: a chatbot talks, automation repeats a fixed recipe, and an agent is the one that gets a goal and figures out the recipe as it goes. The thing that turns a plain chatbot into an agent is exactly two additions: tools and a loop.
03 / 05 · The loop, live
Watch an agent think.
The heart of every agent is one small loop, repeated until the goal is met: Think → Act → Observe. Give it a goal and press play. Each cycle lights up below.
1
Think / plan
Break the goal into the next single action.
2
Act
Call a tool: read_calendar(), search, draft.
3
Observe
Read the result the tool sent back.
↻ repeat until done, or until it hits a checkpoint that needs you
🎯 Goal:Book a 30-min meeting with Sarah next week
Press ▶ Run the agent to watch it work the goal step by step.
🛑 Checkpoint: needs your approval
🔒 100% localThis runs entirely in your browser. Nothing is sent anywhere. The tools and calendar are canned demo data.
💡 Notice the loop didn't blast out the invite. It paused at the action that affects a real person, and asked you. That human-in-the-loop checkpoint is the whole game (more on Day 10).
04 / 05 · When (and when not) to use one
Do you actually need an agent?
Agents are powerful, but they're not always the right tool. The honest rule of thumb: match the tool to how much the task changes.
⚙️ Reach for automation
Cheaper, faster, more predictable.
✓ The task is the same every time
✓ Clear, rule-based, structured input
✓ No judgment call required
› e.g. "When a form is filled in, add a row"
🤖 Reach for an agent
When the path can't be scripted ahead.
✓ Needs judgment on messy input
✓ Branches on what it finds
✓ Stitches several tools together
› e.g. "Sort my inbox and draft replies"
Be honest about the risks
It can loop. Given a vague goal, an agent can keep trying, so you cap its steps and budget.
It can err. It can call the wrong tool or misread a result, same as a new hire on day one would.
It can over-act. The fix isn't "hope it behaves." It's guardrails: read-only by default, and confirm-before-send on anything that touches the outside world.
That confirm-before-send checkpoint (the one the demo just showed) is its own lesson: Day 10 →
05 / 05 · Done · Day 6 of 30
You now understand AI agents better than most people who use them daily.
You can define an agent in one line (a model with a goal, tools, and a loop), tell it apart from a chatbot and from plain automation, and you watched the think→act→observe loop pause for a human at exactly the right moment.
An agent is only as useful as the tools you hand it. See this idea doing real work: