Python, APIs and LLM basics
Everything in this roadmap is built out of one thing: an agent, a language model that calls tools in a loop. Before you build one, and then a whole team of them, you need Python you can read and write, a real grasp of how a model call works over HTTP, and a feel for how a model behaves once tokens and cost come into it. These three short courses cover exactly that. Skip any you already know, but be honest, because every agent you build later assumes all three.
Python for AI
APIs for AI
LLM FundamentalsBuild one agent first
A crew of confused agents is worse than one good one, so you build a single working agent before anything else: a model that decides which tools to call and keeps going until the task is done, with output your code can trust. Everything in the multi-agent stages that follow is built out of an agent shaped exactly like this one.
OpenAI Agents SDK
Pydantic AIAgent workflows and state
A single agent in a loop is fine until you need a branch, a retry, a step that must not run twice, or memory that survives a restart. LangGraph, the framework most teams now reach for, turns an agent into an explicit graph of steps you can see and control, and gives it state that persists. You build agents this way for the rest of the roadmap, single ones here and teams of them in the stages that follow.
LangGraphCrews and role-based teams
The first way to use many agents is to give each one a role, a goal and the tools for its part, then let the team work a task together. This stage builds a crew that way, the people-shaped approach to multi-agent, and is honest about when a readable team of specialists helps and when it only multiplies the ways a system can go wrong.
CrewAIHandoffs and orchestration
A readable crew is one option; explicit control is the other. This stage is about the precise ways agents pass work: one agent handing to a specialist or calling it as a tool, and agents placed inside a graph where the order, the branches and the parallel work are all things you can see. This is how a multi-agent system stays understandable as it grows.
OpenAI Agents SDKLangGraphTools and memory for agent teams
A team of agents is only as useful as the systems it can reach and what it remembers between runs. This stage connects agents to real tools once through MCP, adds managed OAuth so they can act in a user's own accounts, and gives them memory that carries facts from one run into the next.
Model Context Protocol
Composio
Mem0Ship a multi-agent system
A multi-agent system on your laptop is not a product. This final stage puts it behind a real endpoint with a real model, traces every agent so you can see who did what, and deploys the whole thing, so a system of agents becomes something a business can run.
LangGraph
LangfuseAgent engineer
7 stages, 17 frameworks and 5 projects. Finish them and you will have shipped the kind of systems these teams hire for, with the lessons to back every one.
Browse all frameworksQuestions about this roadmap
What does an AI agent engineer do?
They build systems where a language model decides what to do next and calls tools to do it, and they design how several agents divide a task when one prompt is not enough. The work runs from the agent loop and typed output, through crews, handoffs and orchestration, to the tools, memory and deployment a multi-agent system needs.
Can I start this roadmap as a beginner?
Yes. The first stage teaches Python from your first line of code, then APIs and how language models work, and you build a single working agent before any multi-agent work. Every later stage assumes only the stages before it.
Do I need an API key or a paid model account?
No. The courses use small stand-in models you write yourself, or a small open model that runs on your own computer, so no lesson needs an API key or a paid account. Most courses end with a lesson showing the one change that swaps in a real hosted model.
How is this different from the AI Forward Deployed Engineer roadmap?
The forward deployed roadmap is an all-rounder: scope, build, evaluate, deploy and hand over to a customer. This one goes deeper on the agents themselves, and especially on multi-agent systems: crews, handoffs and orchestration, which the forward deployed roadmap only touches in one stage.
Is multi-agent always the right answer?
No, and the roadmap is honest about that. A single well-built agent beats a crew of confused ones. You learn to build one agent properly first, and to split into several only where it makes a system easier to follow.
Do I have to learn every framework on the roadmap?
No. Where two frameworks do the same job, a stop offers both and finishing either one completes it. Knowing one agent SDK, one multi-agent framework and one orchestration tool well matters more than knowing all of them a little.