EECS 498-016 is a new course at the University of Michigan teaching students to build production-grade agentic systems from scratch — grounded in the architecture of OpenClaw, a real open-source multi-agent platform.
You won't just use AI tools. You'll understand the fundamental patterns behind production agent platforms — agent loops, tool registries, session routing, multi-channel message gateways. By the end, you'll have designed, implemented, and defended your own orchestrator.
No toy demos. No prompt-engineering-only courses. Real systems, real code, real engineering.
Course Snapshot
Full technical elective with hands-on projects. Prerequisites: EECS 281, EECS 201 or ULCS, or instructor permission.
Lectures twice a week plus weekly lab sections. Format is all projects, demos, and oral defenses — no exams.
University of Michigan. Applications open now — join the interest list to get updates.
Three Phases, One Goal
Weeks 1–3 · 25% of grade. Master principled AI-assisted coding through Context, Model, and Prompt. Build real software with AI assistance using Aider.
Weeks 4–6 · 25% of grade. Build multi-agent communication protocols from scratch — isolated state, structured handoff protocols, pluggable LLM backends.
Weeks 7–14 · 50% of grade. Design and implement a production-grade orchestrator with task systems, tool registries, agent delegation, and evaluation. Defend your work in an oral exam.
What You'll Build
Projects demonstrating effective AI-assisted development using Aider. Document your process and develop repeatable techniques.
A working system coordinating multiple AI agents with file-based state management, structured protocols, and pluggable LLM backends.
Your capstone: a complete agent orchestration platform with task management, tool execution, delegation, evaluation, and 8+ custom capabilities.
A technical report analyzing your system's performance with empirical data, comparative benchmarks, and design tradeoff analysis.
Present and defend your final project to faculty. Demonstrate technical mastery and explain design decisions under questioning.
Beyond code: a mental model for agentic systems. How agents coordinate, when to delegate, how to evaluate, and what makes systems production-ready.
Grounded in a Real Platform
Most courses teach agent concepts in a vacuum. EECS 498 is grounded in OpenClaw — a real open-source multi-agent gateway with session management, tool execution, plugin architecture, and multi-channel routing. Every pattern you learn maps directly to how agent platforms actually work.
Latest from the Course
Meet the Instructors
Meet the team
