AI Agents: Building Your Minions

Chatbots are so 2023.

You ask a question, it gives an answer. Boring.

What if you could give it a goal?

“Hey AI, plan my vacation to Japan. Book the flights, find hotels under $200, and make sure there’s a Ramen shop nearby.”

This is the promise of AI Agents.

The Loop

Agents work in a loop:

1. Reason: “Okay, I need to book a flight. What do I need to do first?”
2. Act: Call the Expedia API search tool.
3. Observe: “Found 3 flights. Now I need to check hotels.”
4. Repeat: Until the goal is met.

Building One (and failing)

I tried building an agent to organize my email.

Me: “Organize my inbox.” Agent: “Sure!” Deletes all emails. Me: “Wait no.”

Okay, that didn’t happen (thankfully). But the challenge with agents is control.

Tools of the Trade

If you want to build these minions, you need frameworks.

• LangChain: The OG. A bit heavy, but powerful.
• LangGraph: The new cool kid. Think of your agent as a graph of state machines. Much easier to debug.
• AutoGPT: The chaos monkey. Fun to watch, dangerous to run.

The Future

We are moving from “Chat with specific document” to “Go do this thing for me.”

In a few years, we won’t be writing code. We’ll be writing prompt chains for our army of agents.

“Agent, write the rest of this blog post.”

Agent: No.

Well, worth a shot.