Build AI Agents with a Structured Roadmap

What this is, in plain English

This is a free, curated learning roadmap for building AI agents, from the community project “Agent Learning Hub”. An AI agent is a program that runs in a loop, it observes your request, thinks about what to do, and acts (often by using a tool like search or a calculator), repeating until the job is done.

Be honest about what this is: a study path, not a single project you copy and run. There are dozens of agent frameworks and it is easy to drown in them. The roadmap’s value is the order: it tells you what to learn first, what to build at each stage, and which readings actually matter, so you end up understanding agents instead of collecting bookmarks. It is aimed at people ready to write code, so it is advanced.

What you can use it for

• Learn to build agents methodically. Go from “what even is an agent” to a working one without getting lost.
• Build a research assistant. An agent that searches, summarizes papers, and cites sources, great for a final-year project.
• Automate support triage. An agent that reads incoming questions and routes the hard ones to a human.
• Query your own data. An agent that pulls from a database and drafts a quick report.
• A syllabus for a study group. A shared, sensible path for several people learning agents together.

Tools you need

• LlamaIndex and LangChain (free): popular open-source frameworks for building agents and connecting them to data.
• Gemini API (freemium): has a free tier you can learn and experiment on.
• OpenAI and Claude Tool Use (paid): the OpenAI and Anthropic APIs, used for function calling (letting models use tools). They need credits.
• Claude Code (paid): an AI coding assistant that can help you write the agent code.

How it actually works

The roadmap moves through four stages. Work them in order:

1. Learn the core concepts. Understand the difference between a chatbot, a fixed workflow, and a true agent; learn the observe-think-act loop; and learn when an agent is the wrong tool. Recommended reading: Anthropic’s “Building effective agents” and OpenAI’s “A practical guide to building agents”.
2. Build a minimal agent loop. Write a small agent (50 to 150 lines) that: asks the model for structured output (JSON), defines a couple of tools (like search or a calculator), reads the model’s tool requests, runs the tool, and feeds the result back, with basic error handling and a step limit. Use any model’s “function calling” docs (OpenAI, Gemini, or Claude).
3. Add retrieval, tools, and memory. Learn RAG (Retrieval-Augmented Generation: chunk documents, turn them into embeddings, retrieve the relevant bits, and answer with citations). Add real tools (search, databases, files) and memory (short-term, session, long-term), and learn to handle tool failures and made-up answers.
4. Study one real agent system deeply. Pick a mature agent (Claude Code, OpenClaw, and others are suggested) and study how it manages tools, context, permissions, state, and logging, not just the API calls.

Words you’ll see, explained

• AI agent: a program that loops, observing, thinking, and acting with tools, until a task is done.
• Observe-think-act loop: the agent’s core cycle of reading input, deciding, and doing.
• Tool use / function calling: letting the model trigger real actions (search, math, database queries).
• RAG: Retrieval-Augmented Generation, fetching relevant documents first, then answering from them.
• Embedding: turning text into number-lists that capture meaning, so similar things can be found.
• Memory: what the agent remembers, within a reply, across a session, or long term.
• Harness: the surrounding system that manages an agent’s tools, context, and safety.

Original source

Based on the “Agent Learning Hub” GitHub repository by datawhalechina, which offers a structured path and curated resources for learning to build AI agents rather than a loose pile of links.

Notes & variations

• Do you even need a framework? Not at first. The minimal agent loop in stage 2 is about a hundred lines of plain code, and building it teaches more than any framework. Add LangChain or LlamaIndex once you understand the basics.
• Cost: LlamaIndex, LangChain, and the Gemini free tier let you learn for free. The OpenAI and Anthropic APIs and Claude Code are paid; lean on the free options.
• Common pitfall: framework overwhelm, jumping between tools instead of following the path. Finish each stage before moving on.
• Tip for better results: build the minimal loop first. Getting one small agent to pick and use a tool gives you the foundation (and confidence) for everything after.