Build your first AI agent in 2 hours.

There's a strange thing happening with AI agents in 2026. Half the internet acts like building one requires a PhD. The other half thinks you can buy one off the shelf. Both halves are wrong.

An agent is just a loop. A prompt goes in. A model decides what to do. A tool runs. The result comes back. The model decides again. Repeat until done. That's it.

The reason most people never build their first one isn't difficulty. It's that they pick a problem that's too big. "An agent that handles all my email" is a six-month project. "An agent that pulls out every meeting time mentioned in an email and adds it to my calendar" is a Saturday.

The 4 ingredients

You need exactly four things and they all fit on a napkin.

1ModelClaude / GPT, any will do 2-3ToolsRead, write, fetch 1LoopWhile not done, act 1Stop ruleMax steps + done signal

A model is the brain. Tools are the hands. The loop is the body that keeps moving. The stop rule is the thing that saves you from a runaway $40 API bill.

A 2-hour build plan

Saturday morning. Cup of coffee. Open your editor.

Hour one is plumbing. Get an API key. Write a script that sends a prompt and prints the response. Add a tool — a Python function the model can call by name (read_file, web_search, whatever your problem needs). Run it once with a simple task. Confirm the model picks the tool, the tool runs, the result comes back. That's the whole loop, in primitive form.

Hour two is the use case. Pick one thing you do every week and would happily delegate. Not "all my email". One thing. Examples that fit:

• "Read my last 20 transactions and group them by category"
• "Watch a folder, when a CSV lands, summarize what changed"
• "Given this PDF, pull out the three numbers I care about"

Build the smallest version. Skip auth. Hardcode the inputs. Print to stdout. Done.

Try this · Saturday

The 2-hour starter exercise

Pick one task you do every week, takes you 15 minutes, and you'd happily delegate. Build the dumbest possible agent that does it. Hardcoded inputs are fine. Print to stdout is fine. Just make the loop run once, end to end.

First-agent rules of thumb

What goes wrong

Three failures, all preventable.

Vague tools. A tool called do_stuff will be called for everything and accomplish nothing. Name your tools like API endpoints. extract_dates_from_email(text) → list[date]. The model picks tools partly by reading their name.

No stop rule. Loops will happily run forever. Set a max-step count of 5 or 10 on your first build. You can raise it later.

Premature framework. You don't need LangChain. You don't need a vector database. You don't need an orchestrator. You need a while loop. Add a framework when you've felt the pain it solves, not before.

What you'll learn

By Saturday afternoon you'll know something most people who post about agents on LinkedIn don't: agents are not magic. They're a loop and four pieces. Once you've built one, the AI engineering blog posts about "advanced agent architectures" suddenly make sense — they're describing the same loop with more pieces bolted on.

Start with the loop. Add complexity later, and only when it's earned.