Explainer

How AI Agents
Actually Work.

If you have ever wondered what makes an "AI agent" different from ChatGPT, this is the page. Plain English. Receipts where they exist. No hype.

The one-sentence version

An agent is a program that takes a goal, decides which tools to use, performs work in the real world, watches what happens, and adjusts. A chatbot only produces text. An agent does things.

Side-by-side

Chatbot (like ChatGPT before tool use)

You type a question. It produces an answer. The interaction ends when you close the tab. Nothing has changed in your files, your calendar, or your codebase.

Agent (like Claude Code or Codex)

You describe a goal. It opens your files, runs commands, calls APIs, writes new code, deploys it, then reports back. The work persists after you close the tab.

The agent loop

Every modern agent runs the same 4-step loop. Once you see it, you cannot unsee it.

01

Reason

The model reads the current goal, the chat history, and the list of tools available. It decides what to do next. If the goal is "ship a landing page," the next step might be "read the existing index.html to learn the style."

02

Act

The model calls a tool. Tools are functions like Read, Write, Bash, Search, Fetch. Each tool returns structured data (file contents, command output, search results).

03

Observe

The tool's output gets fed back to the model. If a command failed, the model sees the stderr. If a file did not exist, the model sees the error. The model now knows more than it did one step ago.

04

Decide

The model picks the next action. Loop back to step 1. The agent keeps looping until the goal is met, you cancel it, or it asks for permission to do something risky.

The whole magic is in the loop. The model is not magic; the loop is. Give a competent model a small set of useful tools and a clear goal, and it can ship real software.

What tools an agent typically has

This varies by setup, but a productive coding agent usually has these. We teach each one in workshops linked at the bottom.

File ops

Read, Write, Edit

View and modify any file in the project

Shell

Bash

Run commands like npm test, vercel --prod, ls

Search

Grep, Glob

Find code by content or filename pattern

Web

Fetch, Search

Pull docs, check API status, research

External tools

MCP servers

Connect Gmail, Notion, Stripe, your DB

Sub-agents

Task

Spawn a focused sub-agent for one job

Why this matters for non-coders

Traditional software development requires a human to know which file to open, which API to call, which command to run, in which order. The skill ladder is long: syntax, then frameworks, then ops, then deployment.

An agent compresses that ladder. You describe the outcome. The agent already knows the syntax, the frameworks, the deploy commands. What you bring is judgment: what to build, what good looks like, when to stop.

That is the operator role. It is not coding. It is also not "no code." It is something new, and it pays.

Why this matters for coders too

If you already code, an agent does not put you out of a job. It changes what your job is. Instead of writing every line, you direct the agent to write the lines, then review them at the diff level. Speed goes up. Errors go down. The work feels less like typing and more like reviewing a junior engineer who is fast, careful, and never tired.

What you actually need to learn

Three things, in this order:

  1. How to describe outcomes precisely. Vague prompts produce vague output. We teach this in Workshop 02 and the prompt library.
  2. How to configure the agent for your work. CLAUDE.md, AGENTS.md, and MCP server setup. Templates and patterns are in the docs.
  3. How to verify what the agent built. Tests, reviews, deploys. Workshops 06 through 12 cover the verification loop.

What this is not

Ready to direct one?

Start with a $150 build call.

30 minutes. You bring the task. We ship your first agent live, on your screen, in your environment. Credit applied to any tier if you continue.