EECS 498-016: Applied Agentic Software Engineering (University of Michigan, Fall 2026)
by University of Michigan
Michigan's public slides and demos for learning to drive a coding agent, then taking one apart.
Overview
EECS 498-016, Applied Agentic Software Engineering, is a four-credit University of Michigan technical elective running from August 31 to December 11, 2026, taught by Marcus Darden. The course has three graded phases: Apply, Analyze and Create. In Apply, students use Aider with local models served by Ollama or llama.cpp, working through 16 guided lessons and a pair-programmer build. Early lectures cover how LLMs write code, the Big Three framework of context, model and prompt, and spec-driven development. In Analyze, students extend Aider with a skills loader, a model-driven loop, shell tool approvals and a stdio MCP client, while lectures cover messages and streams, stop conditions, retries, approval gates, skills and the Elephant-Goldfish model, shell and browser tools, and subagents. Create builds on the Codex app-server, with lectures so far on compaction, prompt caching, agent memory and knowledge bases. Later lectures were still marked TBD in the syllabus when checked on November 5, 2026. Three supervised hackathons replace exams. Outsiders can use the public materials repository, which holds slides as PDF, PPTX and Slidev source, one demo codebase per lecture, and the Elephant-Goldfish skills. Lecture audio episodes are also published. Assignment repositories are private, and the 16 Aider lessons are not linked publicly, while the course cannot be audited, so you get the content without grading or staff support.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Course
- Cost
- Free
- Duration
- 15 weeks at about 12-15h/week for enrolled students (Aug 31 to Dec 11, 2026); the public slides and demos are self-paced
- Provider
- University of Michigan
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Diagnose AI coding failures using the course's Big Three framework of context, model and prompt
- ✓Drive a coding agent deliberately with spec-driven development while being able to explain every submitted line
- ✓Explain how a production coding agent handles skills, tools, approvals, its loop, context, memory and models
- ✓Implement a skills loader, model-driven loop, shell tool approvals and a stdio MCP client on Aider
- ✓Handle stop conditions, errors, retries and approval gates in the control loop of an agent
- ✓Use subagents, compaction, prompt caching, agent memory and knowledge bases to manage long agent sessions
- ✓Build on the Codex app-server and judge when to write your own loop instead
- ✓Run a written design review in a fresh context before implementing changes to a real codebase
Highlights
- •Students take a real open-source tool apart: the Analyze project adds a skills loader, approvals and an MCP client to Aider
- •The public materials repository had 333 GitHub stars and 50 forks on November 5, 2026, and was updated on November 3
- •Runs on a laptop at $0: local Qwen models through Ollama or llama.cpp, with no GPU or paid LLM subscription required
- •Packages the Elephant-Goldfish design and review workflow as five installable skills in the course repository
- •AI use is required and understanding is graded, with supervised hackathons in place of exams
Who It’s For
Best For
- ✓Working programmers moving from casual AI autocomplete to deliberate use of coding agents
- ✓Engineers who want to understand harness internals before building their own agent loop
- ✓Instructors and team leads looking for a reusable agentic software engineering syllabus and slides
Prerequisites
- •Strong Python, shell and git skills; enrolled students need EECS 281 and EECS 201, upper-level CS standing, or instructor permission
- •No prior AI coding experience or machine learning background is required
- •A laptop that can serve a small local model; no GPU is required
FAQ
What is EECS 498-016: Applied Agentic Software Engineering (University of Michigan, Fall 2026)?
EECS 498-016 is a University of Michigan upper-level course for programmers learning to build software with coding agents using Aider and Codex. Its public slides and demos cover context, specs, tool calls, approvals, subagents, compaction and memory, so you can work out how a production harness behaves and extend one yourself.
Is EECS 498-016: Applied Agentic Software Engineering (University of Michigan, Fall 2026) free?
EECS 498-016: Applied Agentic Software Engineering (University of Michigan, Fall 2026) is free to access.
What level is EECS 498-016: Applied Agentic Software Engineering (University of Michigan, Fall 2026) for?
EECS 498-016: Applied Agentic Software Engineering (University of Michigan, Fall 2026) is aimed at a intermediate audience. Recommended background: Strong Python, shell and git skills; enrolled students need EECS 281 and EECS 201, upper-level CS standing, or instructor permission, No prior AI coding experience or machine learning background is required, A laptop that can serve a small local model; no GPU is required.
How long does EECS 498-016: Applied Agentic Software Engineering (University of Michigan, Fall 2026) take?
Expect roughly 15 weeks at about 12-15h/week for enrolled students (Aug 31 to Dec 11, 2026); the public slides and demos are self-paced. Most learners work through it at their own pace.
What will I learn from EECS 498-016: Applied Agentic Software Engineering (University of Michigan, Fall 2026)?
You'll learn: Diagnose AI coding failures using the course's Big Three framework of context, model and prompt; Drive a coding agent deliberately with spec-driven development while being able to explain every submitted line; Explain how a production coding agent handles skills, tools, approvals, its loop, context, memory and models; Implement a skills loader, model-driven loop, shell tool approvals and a stdio MCP client on Aider; Handle stop conditions, errors, retries and approval gates in the control loop of an agent; Use subagents, compaction, prompt caching, agent memory and knowledge bases to manage long agent sessions; Build on the Codex app-server and judge when to write your own loop instead; Run a written design review in a fresh context before implementing changes to a real codebase.
Topics
Sources
This page was written from 5 sources, 2 on domains other than eecs498-aase.github.io.