Start here
I don't know where to start
Open any AI engineering newsletter and you'll meet thirty words in a week: agents, MCP, Skills, RAG, embeddings, evals, harnesses, guardrails. It looks like thirty separate things to learn. It's really five layers of one system, and they're easiest to learn in a set order.
This page gives you the map and the order. It takes about ten minutes to read. After that, the roadmap will take you through it week by week.
What is the field made of?
Think of an AI app as five layers, stacked one on top of the other. Each layer needs the one below it. Every term you hear belongs to one of these layers.
-
1
Capabilities: what the agent can do
The model itself, the instructions you give it, and the tools it can call. An agent is a model that decides for itself which tool to use next.
models · prompting · structured outputs · tool calling · Skills · MCP -
2
Knowledge and context: what it knows
A model only knows its training data and whatever you put in front of it. This layer is about choosing what goes into that limited space, and fetching it from your own documents.
RAG · embeddings · reranking · context engineering · memory · knowledge graphs -
3
Execution and control: how it works through a task
The loop that runs the model, carries out its tool calls, recovers from errors and knows when to stop or ask a human.
workflows vs agents · orchestration · state · harnesses · human-in-the-loop -
4
Quality and trust: how you know it works
Tests for systems whose answers vary from run to run, records of what the agent did and why, and limits on what it's allowed to do.
evals · tracing and observability · guardrails · prompt-injection defence -
5
Production: how you run it for real
Keeping it fast, cheap, secure and up when real people depend on it.
deployment · caching · cost and latency · retries and fallbacks · governance
Recommended learning priority
If your goal is to become proficient in AI Engineering, I would prioritize the concepts in this order:
-
Start with the fundamentals: LLM APIs, tool calling, structured outputs, prompt engineering, and tokens.
-
Learn how agents work: Skills, MCP, context engineering, memory, orchestration, and state management.
-
Learn how agents use knowledge: RAG, embeddings, semantic search, reranking, and knowledge graphs.
-
Learn how to make agents dependable: Harness engineering, guardrails, human approval, evals, and observability.
-
Learn production engineering: LLMOps, security, reliability, latency, caching, and cost optimization.
My recommendation: Don't treat these as isolated buzzwords to memorize. Build one small agent and use it to understand how the concepts interact. For example, an AI testing agent could use Skills to load testing procedures, MCP to access tools, RAG to retrieve previous test cases, orchestration to execute the workflow, evals to measure quality, and observability to diagnose failures.