도우미에서 에이전트로

So far you've used AI like a chat partner: you ask, it answers, you copy the code. That's an assistant.

An agent is different. You give it a goal, and it acts on its own: it reads files, runs commands, checks results, and keeps going until the job is done. You're no longer the one pasting everything.

This lesson is about that leap: from AI that answers to AI that does the work.

The simplest way to feel the difference:

• Assistant: "Here's the code for a login form" — you take it from there.
• Agent: "I created login.html, wired up the form, ran the app, and it works." — it did the steps.

Tools like Claude Code and Cursor's agent mode are agentic: they touch your real project files, not just a chat box.

Under the hood, an agent runs a simple loop: think → act → observe → repeat.

It plans a step, takes an action (edit a file, run a command), looks at what happened, then decides the next step. It keeps looping until your goal is met or it gets stuck.

Here's that loop sketched as JavaScript pseudo-code so you can picture it:

You don't micromanage an agent. You hand it an outcome and let it figure out the steps.

Compare a single instruction to an agent goal: the goal describes the finished result, not each keystroke.

The big shift: an agent works inside your actual files. It can open index.html, find your CSS, add new code, and save — all without you copy-pasting.

That's powerful and a little scary. It means the agent can change a lot at once, so you'll learn to review what it did (and you'll cover safety in the last lesson of this course).

Beyond editing files, agents can run commands: start your app, install a package, or run tests. Then they read the output and react.

If a command fails, a good agent notices the error and fixes it — no need for you to copy the red text into chat. This is why agents feel like a teammate, not a tool.

When you run an agent in Claude Code or Cursor, you'll see it narrate its steps: "Reading file... editing... running... done."

This is your window into its thinking. Read it! If it heads in the wrong direction, you can stop it and steer. You stay the pilot; the agent just flies the plane for the boring parts.

Most agent tools ask for permission before doing something risky — like deleting files or running an unknown command. They show you the action and wait for a yes.

Treat every approval as a real decision. "Allow once" for normal edits is fine; pause and read carefully before approving anything that removes or overwrites a lot.

Agents shine on tasks with many small, mechanical steps:

• Renaming something across many files
• Adding a feature that touches several places
• Fixing a bug, running the app, confirming it's gone
• Setting up a new project from scratch

For a one-line question ("what does this function do?"), a plain assistant chat is faster. Match the mode to the job.

Here's the mindset to carry forward: you own the goal and the judgment; the agent owns the legwork.

You decide what success looks like and check the result. The agent handles the how — the dozens of little steps in between. That partnership is the heart of vibe coding at the next level.

Before the quiz, picture it: instead of asking "how do I add a contact form?" and copying code, you say "Add a working contact form to my site and confirm it submits."

Same idea, totally different workflow. The agent reads your project, builds it, tests it, and reports back. You review and approve. That's agentic building.

You've made the leap in mindset:

• An assistant answers; an agent acts on a goal.
• Agents run a think → act → observe loop on your real files.
• You hand over a goal (and a "done when"), not every step.
• Watch and approve — you stay the pilot.
• Tools like Claude Code and Cursor agent mode make this real.

Next up: giving your agent tools so it can reach the web, files, and code.