AI Agents from Scratch: How They Work & How to Build One

👁 126 reads

An AI agent is a program in which a language model runs in a loop, choosing a tool, examining the result, and deciding what to do next until it completes the task. To build your first agent and truly understand what’s happening, go through three stages: model APIs, automation, and agents. This page is a roadmap for that path, a 30-day learning plan, code I ran on my own computer on October 4, 2026, and a list of the best free courses and channels I checked the same day.

Roadmap: From API to Your Own Agent

The most common mistake at the start is jumping straight into an agent framework. It works until something breaks, and then you don’t know where to look. That’s why the path has three stages, each building on the previous one:

Stage 1 is covered by a separate guide, AI API. Here I focus on stages 2 and 3. If you prefer clicking to writing code, you can also go through stage 3 in Gumloop or in the AI Agent node in n8n, but understanding will come faster when you see the loop once in 76 lines of Python.

How an Agent Works: The Loop in Four Moves

Every agent, from my simplest one to Claude Code, does the same thing. The model doesn’t execute anything itself: it only requests a tool, and your program executes it and returns the result.

In code it looks like this (fragment of my agent, full version in the article about building an agent):

    for step in range(1, MAX_STEPS + 1):
        r = client.chat.completions.create(model=MODEL, messages=messages, tools=TOOLS)
        cost += getattr(r.usage, "cost", 0) or 0
        msg = r.choices[0].message
        messages.append(msg.model_dump(exclude_none=True))
        if not msg.tool_calls:
            print(f"--- finished after {step} steps, cost {cost:.5f} USD")
            return msg.content
        if msg.content: print(f"[step {step}] THOUGHT: {msg.content.strip()[:300]}")
        for call in msg.tool_calls:
            args = json.loads(call.function.arguments or "{}")
            try:
                obs = TOOLS_DICT[call.function.name](**args)
            except Exception as e:
                obs = f"ERROR: {e}"
            print(f"[step {step}] ACTION: {call.function.name}({json.dumps(args, ensure_ascii=False)})")
            print(f"[step {step}] OBSERVATION: {obs[:160]!r}")
            log.write(json.dumps({"t": time.strftime("%H:%M:%S"), "step": step, "tool": call.function.name,
                                  "args": args, "observation": obs[:500]}, ensure_ascii=False) + "n")
            messages.append({"role": "tool", "tool_call_id": call.id, "content": obs})
    return f"Stopped: limit of {MAX_STEPS} steps (cost {cost:.5f} USD)."

Three things worth remembering:

• Tools are ordinary functions with a description in JSON Schema. The mechanism for requesting a function is function calling.
• Each step sends the entire history, so cost grows faster than the number of steps. That’s why I trim tool results.
• An agent without a step limit and log is asking for trouble. In my code the limit is 8 steps, and every decision goes to a JSONL file.

30-Day Learning Plan

The plan assumes one hour per day and Python basics (variables, functions, for loop). If you don’t have them, add a week for a free Python course before day 1.

What I Built and How Much It Cost

Instead of describing theory, I ran each stage on my own computer. Model for everything except Claude Agent SDK: openai/gpt-5.6-luna via OpenRouter (0.20 USD per million input tokens). I computed embeddings locally in Ollama on RTX 4090.

Frameworks and Tools: What to Choose

Detailed measurement of five libraries (lines of code, time, cost, dependencies) is in the article AI Agent Frameworks. Hosted agents from OpenAI are described in Agents API from OpenAI, and ready-made coding agents in AI Agents for Code.

Testing in Polish

All my tests were in Polish: commands, tool descriptions, documents, and responses. Large models handle this without problems. Three places where Polish makes a difference:

• Document search: embedding model must know Polish. On Polish laws bge-m3 found the right provision in top five in 8 out of 10 cases, and nomic-embed-text only in 4 (comparison).
• Local knowledge: holidays, NBP exchange rates, regulations. Model probably doesn’t know them, so I provide them as tools.
• Function names: without Polish characters, because APIs only allow Latin letters, digits, and underscores in them.

Where to Learn More: Free Courses and Channels (checked Oct 4, 2026)

This is the answer to a question I got from a reader: where do they explain step by step how APIs work, how to do automation and agents, so you really understand. I opened every address on October 4, 2026. For each I write who it’s for.

Official Courses and Documentation

In Polish

• AI Agents for Beginners (Microsoft, Polish version): best free agent course with Polish text, from concepts through tool use and MCP to security. Machine translation, video in English.
• Kacper Sieradziński (YouTube): Python, automation, n8n, Claude Code and agents in business, about 28.5k subscribers. On Dokodu blog step by step agent on your own data in Google ADK. For beginners and intermediate.
• overment, Adam Gospodarczyk (YouTube): recordings of AI_devs livestreams about designing tools for models and agents, about 30.7k subscribers. For programmers, very technical.
• SekurakTV: Building your own AI agent!: recording of free training from September 14, 2026, from security perspective. I verified title and description, didn’t watch the whole thing.
• JSystems: AI agent step by step: same agent shown twice, in n8n and in Python with function calling. Good for understanding what n8n does under the hood.
• cyberFolks: free n8n course: 10 video lessons, over 60 minutes, for complete beginners without programming (requires email).
• Przeprogramowani (YouTube): programmer’s work with AI agents in code, about 20.9k subscribers.
• Karol Kozłecki (YouTube): practical n8n and Make tutorials for entrepreneurs, less about API mechanics.
• unknowNews: weekly link roundup by Jakub Mrugalski, regularly with agent materials. Not a course, but good source to stay current.
• AI_devs: best-known Polish course on building agents for programmers. Paid, registration announced for fall 2026; livestreams on overment channel are free.

English Channels

• freeCodeCamp: full, multi-hour courses for free, including “AI Agents For Beginners” (3 h, July 2026) and “Learn Python and Build Autonomous Agents” (6 h).
• Dave Ebbelaar: building AI systems in Python from engineering side (RAG, FastAPI, automation).
• LangChain: short concept explanations and LangGraph demos, new videos almost daily.
• Sam Witteveen: quick, technical reviews of new models and agent frameworks with code.
• Anthropic and OpenAI: official launches and conference recordings, fewer step-by-step courses.

My sequence for someone who wants to understand, not just click: Anthropic article on agents, then articles from this path with code, then Hugging Face Agents Course or Agentic AI by Andrew Ng, and finally course on chosen framework.

ai-agents-starter Package

All code from this path in one archive, with README in Polish and .env.example file without keys:

• 01_function_calling.py: model calls weather, calculator, and Wikipedia;
• 02_agent_bez_frameworka.py: agent in 76 lines with step limit and log;
• 03_agent_openai_sdk.py: same agent on OpenAI Agents SDK;
• 04_rag.py: chat with PDFs, embeddings in Ollama;
• 05_serwer_mcp.py: MCP server with Polish tools;
• 06_webhook.py: webhook from form with AI;
• frameworki/: one agent in 5 versions;
• dane/: fictional invoices and 3 laws from Official Journal of Laws.

Download ai-agents-starter.zip (2.1 MB, 25 files). MIT license. You enter OpenRouter key in your own .env file. Never paste it into code or screenshots.

Frequently Asked Questions

How to build an AI agent step by step?

1. Master API request.
2. Learn function calling.
3. Close it in loop with step limit.
4. Add read-only tools and log.
5. Test on data whose result you know.

The whole process with code is in the article How to Build an AI Agent in Python.

Do I need to know programming to make an AI agent?

No, n8n, Make, and Gumloop let you click. But to understand what’s happening and why the agent makes mistakes, it’s worth reading a few dozen lines of loop code once. You can do this without prior experience.

How much does learning to build agents cost?

Courses from the list are free. For tokens: all my tests from this path cost about 0.15 USD. To start, 2-5 USD top-up with limit on key is enough.

How does an AI agent differ from automation?

In automation you set the steps. In an agent the model chooses the order of steps based on tool results. Most practical implementations are automation with one agent step.

Where to start: n8n or Python?

If you want working automation quickly, from n8n or Make. If you want to understand API and agent mechanics, from Python. Best both: same flow first in code, then in n8n.

What model to choose for first agent?

Cheap model with tool support (on OpenRouter filter “tools”). I used openai/gpt-5.6-luna at 0.20 USD per million input tokens. Take stronger model only when cheaper one clearly loses the plan.

What is MCP and do I need to know it?

MCP is a standard for connecting tools to AI applications. Not needed for first agent, but useful when you want to use your tools in Claude Code, Cursor, or other client. See Your Own MCP Server in 15 Minutes and introduction MCP: What It Is.

Can an AI agent run on my computer without internet?

Yes, with local model in Ollama, which has OpenAI-compatible API. Smaller local models more often confuse tool arguments, so start with simple tasks.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and Verification Date

• All code runs: my computer (Windows 11, Python 3.13, RTX 4090), October 4, 2026.
• Package versions: PyPI and npm, Oct 4, 2026. Make, Zapier, n8n pricing: vendor sites, Oct 4, 2026.
• Courses and channels: every address opened Oct 4, 2026. Subscriber counts rounded by YouTube.
• Earlier articles on blog: AI Agent: What It Is, AI Agent: What It Does and How to Test It, Automation with AI, RAG, Fine Tuning.

Kinetyra