Lately, everyone has been talking about “AI Agents”, a buzzword that seems to promise technological revolutions overnight. This 8-chapter series aims to go beyond the marketing hype and explain, in simple terms, what an autonomous assistant really is and how it actually works in practice.
“Behind the grandiose names used in the online world, there is, in fact, a simple logic. We do not need science-fiction scenarios to understand this technology. We start with the simplest example of a digital assistant and analyze it step by step, to see exactly where the marketing ends and the concrete reality begins.”
To follow this series, you do not need programming knowledge or huge budgets. The structure has been designed so that anyone can follow it:
- The introductory chapters (Episodes 1-3 and 7-8): These can be read independently and do not require any technical knowledge. They are based on logic, everyday examples (such as photo processing or online stores), and simple analogies, without technical terms that might intimidate you.
- The technical chapters (Episodes 4-6): These take things one step further and look a little behind the screen. They are not incomprehensible to an ordinary reader, but they do contain real lines of code and configurations used to put the assistant to work. Reading them carefully is only necessary if you intend to reproduce and build this experiment on your own computer.
For the best visual experience, we recommend viewing the articles on a Desktop computer. We have integrated a series of interactive graphical simulations into the pages to illustrate the assistant’s decision-making flows; due to space and presentation constraints, these simulations are not visible on mobile phones, where following them would be extremely cumbersome.
🔹How the Assistant Works: What We Will Discover Together
The central project in this series revolves around a very simple example: a warehouse manager who checks the stock of components. Instead of allowing a Cloud-based program to have direct access to your computer or your data, we have built a completely separate system in which each component has a clearly defined role.
The system acts as an interface between the two worlds. The artificial intelligence in the Cloud simply receives the user’s question and transforms it into a simple command, but it does not have direct access to your network. A script running on your computer takes that command, checks the product list on the local server, and sends back only the required answer. This way, your private data remains within your own network, without being transmitted to the AI service in the Cloud.
🔹The Roadmap of the 8 Episodes: A Chapter-by-Chapter Guide
The series is divided into eight articles, designed as a logical story, from simple rules to the way a program eventually runs on a server:
- Episode 1: Demolishing the Marketing Jargon. We understand how an AI assistant works and how we moved from the fixed rules of the past to systems capable of interpreting nuances, using a simple analogy from the movie The Matrix.
- Episode 2: Tool Design and the Data Myth. We dismantle the idea that AI holds all the information in the Cloud and show how it actually receives only a list of instructions (the remote control manual) that allows it to ask your computer for help.
- Episode 3: The Need for a Draft (Chain of Thought). We explore how AI models can use an intermediate space to break down a problem into steps before formulating an answer, and why this mechanism can be important for solving more complex tasks.
- Episode 4: The Cloud Lab and the First Connection. We take our first step behind the screen. We generate an access key in a test environment and write a simple Python script so that the model can respond to us through structured commands.
- Episode 5: Local Production and Proxmox Infrastructure. We move the experiment to our own server. We start a virtual machine, install a real inventory list in a local database (MariaDB), and keep the passwords secure within the system.
- Episode 6: Pure Programming and Closing the Loop. We write the final code that brings everything to life. We add a filter that automatically interprets word forms in Romanian (including plurals) and open a live chat window directly in the computer terminal.
- Episode 7: The Black Box Concept. We change our perspective and look at the technology from a distance. We analyze how end users see only a single magic button, using practical examples from photo editing and e-commerce catalogs.
- Episode 8: The Great Debate (RAG vs. Function Calling). We conclude the series by drawing a clear line between two different approaches: systems that search for information and meaning in documents, and systems that use functions to obtain exact data, such as inventory levels.
🧠 Total Control: Why Is This Experiment Important?
The real value of these chapters does not lie in writing complicated code, but in the **security philosophy** behind it. Instead of uploading your company or inventory data to external Cloud servers, you use artificial intelligence simply as a “decision engine”.
Your local program controls access to the data and executes the commands received from the AI, within the limits you have defined. This way, you keep important data within your own infrastructure and limit what the AI service in the Cloud can see or access. It is an interesting compromise between the flexibility of an AI assistant and the control provided by your own computer.
🔹The Steps of the Experiment
To understand and simulate this project on your own, the series will take you through four major logical phases:
Step 1: Understanding the Logic
We discover how modern systems work and how they move from a question expressed in natural language to a concrete action, as well as why they need clear rules to avoid incorrect responses.
Step 2: Safe Testing
Running the first lines of code in an online testing environment (Google Colab), verifying that the model generates exactly the commands we need.
Step 3: Preparing the Local Server
Configuring the system on your own computer and creating the SQL database with the inventory list that the assistant will be able to query.
Step 4: Starting the Direct Chat
Running the final Python program, which receives messages from the terminal, queries the local database, and provides the final answer without human intervention.
🛠️ Are You Ready to Start Reading?
You do not need advanced experience, just curiosity. Every chapter is explained step by step. Read them in order, start with the introductory part, and discover how all the components behind the screen fit together!
👇 Access the Series Chapters Below
⚙️ Appendix: A Plain-English Guide to Key Terms
- Function Calling: A mechanism through which an AI model transforms a request expressed in natural language into a structured command (usually in JSON format) that a program can interpret and execute.
- Self-Hosted: Running programs and databases directly on your own computer or server, without hosting them on another company’s infrastructure.
- Determinism: A way of working in which the same input data produces the same result according to the same rules (such as a search through a list), without approximations or assumptions.
- RAG (Retrieval-Augmented Generation): A method in which an AI model receives relevant information extracted from documents or other data sources and uses it to formulate a response. It is useful, for example, when you want an assistant to answer questions b
🔹 Episode 1: Breaking Through the Marketing Bubble. What Is an AI Agent, Really?
Welcome behind the scenes of technology. In this first chapter, we leave behind the grandiose definitions from Silicon Valley’s glossy brochures and strip away the mystery surrounding the real structure of applied artificial intelligence. We dive straight into the code and logic to see where the advertising ends and reality begins.
🔹 Episode 2: Function Calling — Giving Rigid Tools to a Fuzzy Brain in the Cloud
Discover how AI Agents turn natural-language requests into precise actions. We break down Function Calling, JSON schemas, and the ReAct loop to show how a Cloud-based AI model can reason about a request while your local application executes the actual tools and keeps control of your data.
🔹 Episode 3: Chain of Thought. The Internal Monologue and Digital Draft of an LLM
An agent with only executive arms (ReAct), but no ability to plan its steps, is just an impulsive automaton. In this chapter, we leave corporate definitions behind and open the door to the internal monologue of an LLM. We explore why AI needs to build its reasoning on a hidden “digital sheet of paper” before delivering the final answer, and what Inference-time Compute and Reasoning Tokens really mean.
