How to fly a helicopter in 90 seconds
I wrote this post over a year ago, thought it was too trivial and never shared it. Given that the concepts have remained and ballooned, you might as well know —
AI people love to invent words to describe otherwise pretty mundane things, and the word “Agent” was just another one of them. Nobody knows who came up with it, but at one point in the past, code blocks got renamed to agents almost overnight, and the agentic revolution was born.
Agents are code that are constrained to a particular area of expertise. Most importantly, they are supposed to be able to work autonomously within that area.
Remember The Matrix?
Instead of being busy doing everything in the world, Mr. Smith had only one goal - to maintain order and security within the matrix, and to eliminate the threats. Nobody told Smith what to do, he just did what he thought was right. In the same way AI agents are constrained to doing only a specific subset of tasks, and doing it on their own, really well.
General LLMs are just dumb text generators. They know everything about everything, but they aren’t particular good at any one task. If you want a refresher on the origins of LLMs, it’s here—
Agents, on the other hand, are code loops that use the underlying LLMs. One agent “knows” how to book tickets, another one that can update your calendar, another one writes code, and yet another adds milk to your refrigerator when it runs out.
This is helpful because it makes the agent more focused, yielding better responses to specific questions. Being a focused system, an agent gets to prioritize responses for a particular area. Better results for you, faster computer time for the underlying models. Win-win-dial-tone.
These agents work in three easy steps:
First. Sense the environment. This is where they read your input, and analyze it to, trying to understand what you actually want.
Second. Make a decision about the next steps. Using either predefined steps, or some kind of internal logic on how to behave based on the inputs, the agent pick the next steps, be that to search the internet, to parse your data, call other applications, or anything else like that.
Third. Act on the input and repeat. At this stage, the agent does what it planned in step 2, collects all the data, analyzes it, and then sends it back to step 1 to repeat until the original problem was solved.
So that’s what it is, in short LLM agents turn passive language models into proactive goal-seeking loops that “think”, use tools, and iterate until the job is done.
That’s it, no magic, no moat, just code that is told to do the same thing over and over and over again.
Next time someone talks to you about AI Agents with grandeur and valor, you will know what they are talking about. It’s a bicycle, sold a 747.
— Kirill.