EECS 498-016 · Fall 2026 · University of Michigan
AASE · say it "A two S E"
Learn to build software with AI coding agents, then build the agents themselves. The semester runs in three phases: work the tools well, script your own agents straight against LLM APIs, and ship a production-grade orchestrator as your capstone.
01 The premise
Software engineers increasingly direct systems that write code, instead of typing every line themselves. This course treats that work as an engineering discipline and teaches it like one: with theory, deliberate practice, instrumentation, and grading that rewards understanding over output.
You will spend the whole semester inside real tools. First you learn to pair with AI deliberately, and to diagnose failures using the Big Three: context, model, prompt. Then you drop below the abstraction and write your own agent against a raw LLM API, tool calls, loop and all. The back half of the course is a capstone where you design, evaluate, and deploy an agent orchestrator.
Lectures run live demos. Labs are hands-on working sessions with checkoffs. Every project ships code that runs.
"Using the tools is table stakes. The engineers we want to graduate can explain why an agent failed and fix the system around it."
02 Three phases
Work with AI coding tools as a first-class engineering skill. Aider comes first because it makes context, model, and prompt visible and controllable. Claude Code and Codex arrive in week 3. You learn spec-driven coding and the Big Three diagnostic framework, then ship a project with both tools.
Open the black box. Call an LLM API directly, hand the model tools, and write the agent loop yourself in about a hundred lines. Then customize Claude Code with slash commands, hooks, sub-agents, and MCP servers, and learn to test what you build with evals and guardrails.
Design and ship a production-grade agent orchestrator of your own. State and memory, multi-agent coordination, MCP integration, error handling, cost control, observability, and a real deployment. You demo the system live in week 15 and back every claim with an eval suite.
Grade weights are near-final and get locked in the week 1 syllabus.
03 Outcomes
04 The stack
Everything runs on a normal laptop. No GPU, no cluster account. You bring a machine that can run Python and git; the course supplies guidance on models, providers, and budgets.
05 Staff
Seats will be limited while the course is experimental. Join the interest list and we'll email you when registration details are settled.
