What an AI agent actually is (and what you'll build here)

None required. This is the entry point for the Build AI Agents track. If you have never used Claude or any AI chatbot before, start with Your First Useful Output first.

A chatbot answers your questions. An agent goes and does things. That single difference is the entire foundation of this track.

When you type a question into Claude or ChatGPT and get a reply, that is a chatbot (a system that takes your text, thinks about it, then sends text back). It reads, it responds, it waits for your next message.

An agent does something more. It perceives information from the world (an inbox, a spreadsheet, a calendar), decides what to do about it (summarize, flag, reschedule), and acts on that decision (sends the email, updates the row, moves the meeting). That perceive, decide, act loop is what makes it an agent instead of a chat window.

The loop can repeat: the agent checks the result of its action, decides if it worked, and acts again if it did not.

Why this matters to you: a chatbot needs you in the chair, typing each question. An agent can do a whole job while you are away.

It reads your morning inbox, drafts three replies, and has them waiting when you sit down. It monitors a project tracker and pings you only when a deadline slips. It does the work, not the conversation about the work.

Here is what this track builds, rung by rung:

Free (this funnel): 1. Give it one job. Build a single agent that does one real task end to end. 2. Give it tools. Let it read your email, check your calendar, and search the web so it acts on your real information, not only your words.

Premium (what opens next): 3. Memory. The agent remembers what you told it last week instead of starting cold every session. 4. Connect your tools (MCP). MCP (Model Context Protocol) is the standard way to plug services like Gmail, Notion, and Slack directly into an agent. 5. Skills. Saved instructions the agent can call by name, so you build a workflow once and reuse it. 6. Agent teams. Multiple specialists that hand work to each other (a research agent feeds a writing agent feeds a review agent).

For the deep take on what makes something "agentic" and how AI sessions become agentic workflows, two existing lessons cover the concept in full: What an agent actually is in the Google Antigravity track, and Thinking in Agentic Sessions in the core path. This lesson is the front door, not the full explanation. Its job is to show you what you will build here, not re-teach the theory those lessons already cover.

Open Claude (any surface: claude.ai, the desktop app, or Google Antigravity) and paste this prompt:

"Explain the difference between a chatbot and an AI agent in one paragraph. Use a concrete, everyday example (like managing an inbox or tracking a project). Keep it under 100 words."

Read the response. Then ask a follow-up:

"Now describe the perceive-decide-act loop using that same example. Label each step."

Compare what you get to the definition in this lesson. Does the example fit? Does each step in the loop make sense? If anything feels off, ask Claude to try a different example until one clicks.

A chatbot waits for your question and answers it. An agent checks your inbox every morning (perceive), decides which emails need replies and which can wait (decide), then drafts those replies and queues them for your review (act). The difference: you never had to ask. It did the job on its own.

Perceive: the agent reads the 12 unread emails in your inbox.
Decide: it flags 3 as urgent, marks 7 as low-priority, and identifies 2 that need a drafted reply.
Act: it writes the two reply drafts and moves the low-priority emails to a "later" folder.

What good looks like

• You can explain chatbot vs agent in one sentence without looking at the lesson
• You can name the three steps of the agent loop (perceive, decide, act) and give a real example for each
• You tried at least one follow-up question to pressure-test the definition

Think about one task you do every week that follows a pattern: you check something, make a decision about it, then do something based on that decision. Examples: reviewing a report and flagging issues, scanning job listings and saving good ones, checking your calendar and rescheduling conflicts.

Open Claude and paste this:

"I have a weekly task I want to turn into an agent workflow. The task is: [describe your task in 1-2 sentences]. Break it down into the perceive-decide-act loop. For each step, name exactly what the agent would read, what decision it would make, and what action it would take. Then tell me which step would be hardest to automate today and why."

This is the seed for the agent you will actually build in the next lesson.

What good looks like

• You picked a real task from your own week, not a hypothetical
• Claude broke it into perceive, decide, and act steps that make sense for your task
• You can identify which step is hardest to automate and have a rough sense of why