An AI agent is an LLM in a loop in 20 lines of code: where complexity begins

The phrase "build us an AI agent" now sounds as simple as "build a button," even though the word "agent" in practice hides three different things — a prompt, a chatbot, and an actual agent, explains the author of an article on Habr dedicated to the anatomy of AI agents.

How an Agent Differs From a Prompt and Chatbot

According to the author, confusion around agents arises because three completely different things — a simple prompt, a dialogue chatbot, and a full-fledged agent — are called the same. The difference is fundamental: an agent does not simply answer a question, it acts in a loop, making decisions independently about the next step.

What Is the Anatomy of an Agent

At the core of a minimal AI agent, according to the author, lies an LLM running in a loop — such a construct can be implemented in literally twenty lines of code. This is where the real complexity begins: turning a minimal loop into a reliable system that does not waste budget and solves a real problem.

• The basic architecture of an agent — an LLM in a loop, approximately 20 lines of code
• The author identifies three different concepts confused under the word "agent": prompt, chatbot, agent
• The article outlines five ways to accidentally burn through budget on an agent
• A simple filter is given for checking whether a task really needs an agent

What This Means

Demand for "AI agents" outpaces understanding of what they actually are: while teams confuse a prompt, a chatbot, and an agent, budgets continue to be wasted on development. Clear separation of these concepts and a simple filter of "do we even need an agent" is a practical way to avoid overpaying for a trendy term.