The AI agent engineering curriculum

How the course works and what it is pointing at, what the toolchain is and how an exercise is graded, and the completion object every later module reaches into: choices, finish_reason, tool calls and usage.

No prerequisites · First module in track

What a model actually computes, why tokens and the context window are the constraint that binds first, how a sampler turns a distribution into a token, and why a conversation is a list you own and resend.

Requires: Foundations

The parts of a working prompt, what examples teach that instructions cannot, how to get output a program can branch on, and how to measure a prompt change instead of guessing at it.

Requires: LLM fundamentals

Give a model hands. Write schemas it can read, build the ReAct loop from scratch, answer a whole batch of parallel calls including the ones that failed, and end every run with a reason you can act on.

Requires: Prompt engineering

What a framework adds to the loop you already wrote, built rather than described. A registry that derives schemas, five hook points and a trace, and an agent expressed as data you can validate and diff.

Requires: Tool use and building an agent

Give an agent knowledge it was never trained on. Chunking, embeddings, a retriever allowed to return nothing, the pipeline that grounds an answer in what it found, and the separate problem of remembering the conversation.

Requires: Agent frameworks

Everything the model reads is input, and most of it did not come from you. Tool results and retrieved documents are attacker-controlled text landing in the same array as your system prompt, and the model cannot tell them apart.

Requires: Memory and RAG

A second look is cheap. Making the second look independent of the first is the whole difficulty, and it is what separates a check that catches errors from one that launders them into confidence.

Requires: Agent security

Split one agent into several and route between them, orchestrate workers over a plan that can actually be fanned out, grade the trajectory and not just the answer, and answer the three questions production asks of a pile of trace records.

Requires: Self-correction

Rebuild what a graph runtime gives you: channels with reducers, a super-step scheduler that knows which nodes are stale, and a checkpointer that makes resume, time travel and forking fall out of one dict.

Requires: Agent frameworks

The runtime style frameworks, rebuilt small. A handoff is a tool call that reassigns one variable, a guardrail is a function that can raise before the model is ever called, and structured output is a validator wired back into the prompt.

Requires: State graphs and durable execution

One support agent built end to end: grounded answers from a knowledge base, an escalation contract with a handoff payload worth receiving, and a persona driven harness that fails the agent for escalating too early and for refusing too long.

Requires: Multi-agent systems and production

An orchestrator that fans questions out to workers, a supervisor that writes the brief once, and a verification pass that checks every cited claim against the source text before anything ships. With the accountant that shows what fan out costs.

Requires: Support triage that knows when to escalate