The 3 Layers of AI: From Chatbots to Agents

The Journey to Autonomy

As you progress through this curriculum, you will build increasingly complex AI. A common interview question is to differentiate the three "Layers" of AI architecture. Understanding this early helps you know exactly what you are building.

Layer 1: Generative AI

This is what ChatGPT was at launch. The model is literally just an autocomplete engine predicting the next token over and over until the output is finished.

• Characteristics: Stateless, static, and straightforward.
• How it works: You inject a prompt, it generates text, code, or images. It has absolutely no memory aside from the exact words you fed into its current "Context Window".
• Analogy: A brilliant writer locked in a room who can only write an essay based exactly and entirely on the prompt you slip under the door.

Layer 2: Code Agents

Eventually, engineers realized: what if the AI could run the code it writes? What if it could browse the internet and read the results?

• Characteristics: Environment-aware, dynamic iteration.
• How it works: The AI sits inside a ReAct Loop (Reason -> Act -> Observe). It is given a "Scratchpad" (working memory) and access to APIs or Code Execution environments (like Docker). Instead of just answering purely from its brain, it can write a Python script, execute it, read the error log in the terminal, rewrite the script, and run it again.
• Analogy: The writer in the room is now given a laptop with Python and Google. If they don't know an answer, they search for it first, then write the essay.

Layer 3: Agentic AI (Multi-Agent Systems)

Why stop at one generalized agent? AI Engineering has moved toward creating full autonomous organizations of specialized agents.

• Characteristics: Goal-driven, asynchronous, highly autonomous.
• How it works: Instead of one solitary loop, this is a hierarchical network of multiple specialized agents. You have an Orchestrator that takes a complex goal and decomposes it. It assigns work to specific Executors, whose output is directly scrutinized by Critics. They all share long-term context via a Persistent Memory DB (Vector/Graph databases).
• Analogy: The room is now an entire engineering firm. There is a Senior Product Manager outlining features, a Junior Developer writing code, a QA Engineer reviewing the pull requests, and a permanent file cabinet retaining knowledge of all past projects. They run parallel workflows until the overarching goal is met.