I spent the last year trying to seriously learn how AI agents work. The material is out there. The ReAct paper from Yao et al., the MCP spec, framework docs from LangChain and CrewAI, vendor posts from Anthropic and OpenAI, dozens of YouTube videos, Hacker News threads. All useful. None of it sequenced. I'd end up with twenty browser tabs and no clear order to read them in.
So I built build-agents.dev: one free, self-paced path through the material, from "what is an agent" to "how do I run one in production." This is my first Reddit post about it. Looking for feedback, especially the harsh kind.
What it is
Four tracks, 20 modules, 67 lessons. The whole thing is opinionated about one thing: learn the loop before reaching for a framework.
Track 1: Agent fundamentals. The orchestration loop, tool use, ReAct, memory, context engineering, agentic RAG. Lesson 1 builds an agent loop in about 15 lines of plain Python.
Track 2: Orchestration patterns. When single-agent breaks. Supervisor/worker, hierarchies, sequential pipelines, peer-to-peer, state management, agent safety, metacognition.
Track 3: Building with MCP. Model Context Protocol from the protocol up. Hosts, clients, servers, capability negotiation, transports, the three primitives, multi-server orchestration, auth, production patterns.
Track 4: Production agents. Reliability (failure taxonomy, retries, validation, runtime policy), evaluation (testing non-deterministic systems, building eval datasets), observability (structured logging, tracing, dashboards), deployment (containers, queues, human-in-the-loop).
Every lesson has a slideshow walkthrough you can pace through. About half have an interactive Python sandbox that runs in the browser via Pyodide. No local setup, no API keys required to start. Example code uses the ollama Python package because it's free and works locally, but the patterns are model-agnostic.
Why I made it
- The existing material is scattered. No single source takes you from "what is an agent" to "how do I run one in production" in a coherent order.
- Most agent tutorials start with a framework. You install LangChain or CrewAI, follow ten lines, you have an "agent." But you don't actually understand what's underneath, and the moment something breaks you're stuck reading framework source. Lesson 1 doesn't import a framework. It builds the loop.
- I wanted my own reference. As I was learning this for actual work, I kept rewriting summaries. Turning notes into a curriculum was a small extra step.
What it doesn't do
- Not exhaustive. No track on fine-tuning, vision agents, or RL.
- Lesson content is hand-written by me, not pulled from primary sources. I'm an engineer, not a researcher. If I got something wrong, I want to know.
- Pyodide sandboxes will run any Python you write, but won't call an LLM unless you set up Ollama locally with
OLLAMA_ORIGINS=*. Most exercises are about agent control flow, not real model calls. - Written by one person. Production patterns reflect what I've actually shipped.
How it's free
Free platform, free curriculum, free sandbox compute (Pyodide runs in your browser). No paywall, no signup required to browse. Sign in with GitHub or Google if you want progress saved across sessions.
If the curriculum saves you the twenty browser tabs I had to deal with, you can buy me a coffee. It's how the project stays free and unsponsored. No pressure, but if it's useful and you can spare a few bucks, it genuinely helps.
Things I'd love feedback on
- Is the track ordering right? Does MCP belong where it is, or before orchestration patterns?
- Are there obvious holes? What did I leave out that you'd expect from a serious agents curriculum?
- Are the prerequisites realistic? Right now it's "Python basics, command line, an LLM you can call."
- Did anything feel hand-wavy or wrong? Especially Track 4 where I'm closest to my own bias.
If you read a lesson and have notes, I'll take them seriously. The whole thing is small enough that I can refactor sections quickly.
Link: https://build-agents.dev