Skip to content

The AI Map

Every useful AI system — from a single well-written prompt to a multi-agent platform — can be described with the same layers. This map is the "big picture" for both courses.

Reference architecture of an agentic AI system Interface chat, API, email, human approval screen Orchestration layer planner and router, workflow engine, state, checkpoints, budgets, stop rules Agents Intake, Researcher, Analyst, Writer, Reviewer Harness system prompts, tools (MCP), guardrails, structured outputs, permissions Context & data RAG pipeline, vector store, memory, multilingual ingestion Models model router; large, small and reasoning models Cross-cutting Evals Observability Security Cost
Layer What it does Question it answers Literacy class Engineering class
Interface Where people meet the system: chat, API, email, an approval screen How do people use it? 5 8
Orchestration layer Plans the work, routes it to agents, keeps state, enforces budgets and stop rules Who does what, in what order, and when do we stop? 7 6
Agents Specialized workers — each a model with its own instructions and tools Which specialist handles this step? 7 6
Harness Everything wrapped around a model call: system prompts, tools, guardrails, output formats, permissions What may the model see and do? 3 2, 5
Context & data Getting the right information into the model: retrieval (RAG), memory, document ingestion Does the model know what it needs to know? 4 4
Models The large language models themselves, and choosing which one to use Which brain, at what cost? 1, 2 1, 6
Cross-cutting Evals, observability, security and cost apply to every layer Does it work, can we see it, is it safe, can we afford it? 6, 8 7, 8

Four kinds of "engineering"

The field's vocabulary grew in the same order as the layers above. Each new term answers the question the previous one could not:

  1. Prompt engineering — writing one excellent instruction. "My prompt is great — why is the answer still wrong?" Because the model lacks information…
  2. Context engineering — deciding exactly what information the model sees. "It has the right information — why can't it act?" Because it has no hands…
  3. Harness engineering — the tools, rules and structure around the model. "It can act — how do I know it does the job well, and stops when it should?"
  4. Loop engineering — the agent loop (act, observe, repeat until a stop rule or a human says stop) and the improvement loop (measure with tests, change one thing, measure again).