Chapter 1 · Part 1

Muse vs. a chatbot

In September 2026 Meta released Muse, which it calls a personal AI agent. Its pitch: Muse "doesn't just answer questions, it actually does the work." It can send your email, book travel, sell things you no longer need and argue with a phone company about your bill.

It's easy to file that under "a smarter chatbot", but it's a different kind of program. This course takes Muse apart piece by piece. The same parts show up in nearly every agent you'll meet, including coding agents, browser agents and the assistants built into your phone.

Start with the most important difference. Scroll to give a chatbot and an agent the same request.

A chatbot runs the model once: your words in, an answer out. The work is still yours.

scroll↓

The loop is the whole trick

Under the hood, an agent is a very ordinary program wrapped around a language model:

  1. Plan. The model reads the goal and everything that has happened so far, then decides what to do next.
  2. Act. It picks a tool, such as "search email", "open this web page" or "send a message", and fills in the inputs. The app runs that tool for real.
  3. Observe. The tool's result goes back to the model as new text: "Found 3 bills, the latest is $85."
  4. Repeat until the model decides it's done or needs to ask you.

The model never touches your inbox itself. It only writes text that says which tool to call and with what, and the surrounding program does the rest. That split between a model that decides and a harness that executes will matter a lot in Part 3, because it's where you put the safety checks.

📌Want to build this loop yourself?

The hands-on Build an AI Agent course writes this exact loop in about 30 lines of Python. This course stays at the "how does a product like Muse fit together" level, so no code is required.

What turns a loop into a personal agent

A bare loop plus a few tools is enough for a demo. Muse adds five more things, and each one gets its own chapter:

  • A capable brain. A reasoning model that can see images and think before it acts. (Chapter 2)
  • Hands. Connectors to email, messaging and payments, plus the ability to work a web browser like a person would. (Chapter 3)
  • Memory. It remembers what matters to you, including things you mentioned only once. (Chapter 4)
  • Stamina. It keeps working after you close the app and splits big jobs across helper agents. (Chapter 5)
  • A guard. The security design that keeps all that power from being used against you. (Chapters 6–7)

The first four are what make it useful. The fifth is what makes it safe to switch on.