Agentic AI for Computer Systems Design

Harvard University • Fall 2025
CS249r: Special Topics in Edge Computing

Instructor: Vijay Janapa Reddi
Teaching Assistant: Arya Tschand aryatschand@g.harvard.edu
Course Staff: Zishen Wan and Chenyu Wang
Time: Monday/Wednesday 2:15-3:30 • Location: SEC. Room 6.412
Office Hours: Monday 4-5pm (Instructor), Wednesday 4-5pm (TA hours) Canvas: Course Site

Course Goal

For decades, computer systems have been meticulously designed by human experts using intuition, heuristics, and manual optimization. Now we’re witnessing a fundamental paradigm shift: agentic AI systems are transforming how we design these systems, from software optimization to chip layout.

This transformation spans the entire computing stack—from AI agents that optimize compiler passes and generate high-performance code, to reinforcement learning agents that automatically design processor microarchitectures, to neural agents that solve chip placement and circuit synthesis. We’re moving from human-designed heuristics to autonomous AI agents that can explore solution spaces too large for manual analysis.

What you’ll learn:

How agentic AI systems are being applied across the complete computing stack
Critical analysis of agentic design methodologies through systematic literature review
Research synthesis skills by writing comprehensive survey papers on cutting-edge topics
How to identify research gaps and propose future directions in agentic AI for computer systems design
Evaluation frameworks for agentic AI in software, architecture, and chip design
Collaborative research skills through team-based survey projects

What is Computer Systems Design?

Computer Systems Design spans the complete computing stack - from software algorithms to physical silicon implementation. It encompasses three critical layers:

Software Systems: Compilers, runtime systems, and algorithms that determine what computations to perform and how to optimize them
Computer Architecture: Processors, memory hierarchies, and system organization that determine how computations are executed
Chip Design: Physical layout, circuit synthesis, and manufacturing processes that implement systems in silicon

Traditional systems design relies on human expertise, heuristics, and manual optimization across vast design spaces with millions of possible configurations at each layer.

In this course, we explore the complete computing stack:

AI for Software: AI systems understand what needs to be computed efficiently
AI for Architecture: AI agents design how to compute it efficiently in hardware
AI for Chip Design: AI tools implement the architecture physically in silicon

What is Architecture 2.0?

Architecture 2.0 is the paradigm shift where agentic AI systems automatically explore, evaluate, and optimize design spaces across the entire computing stack. These agentic approaches leverage reinforcement learning, neural networks, and Bayesian optimization to:

Generate and optimize code by learning from millions of programs and performance patterns
Design compilers that automatically discover better optimization strategies than human experts
Configure processors by exploring cache sizes, pipeline depths, and execution unit combinations
Create accelerators that co-optimize hardware architecture and algorithm mappings
Place and route circuits by finding optimal chip layouts faster than traditional tools
Verify designs by generating test cases and formal proofs automatically

The goal is to create agentic AI systems that can design better computing systems than human experts across software, architecture, and chip implementation - while exploring design spaces too large for manual analysis.

Course Overview

This course explores the shift from human-designed heuristics to agentic AI systems. While conventional courses teach you how existing systems work, this seminar explores how agentic AI will design tomorrow’s complete computing stack.

This course is focused on exploring how agentic AI enables the specialized, efficient systems that edge computing demands. We’ll work hands-on with cutting edge research tools including CompilerGym (software), ArchGym (architecture), and DREAMPlace (chip design).

Students will systematically explore agentic AI applications across the complete computing stack - from code generation and compiler optimization, through processor and accelerator design, to chip placement and verification - while identifying the most promising research directions for agentic computer systems design.

Note: This is a research intensive seminar with limited enrollment, focused on understanding how agentic AI systems design complete computing systems rather than analyzing existing implementations.

Prerequisites & Expectations

Required Background:

Computer Systems: Must have completed courses in computer architecture, compilers, or systems programming. Strong foundation in at least one area: processor design, compiler optimization, or software systems is essential.
Machine Learning: Familiarity with reinforcement learning, neural networks, and optimization methods
Programming: Proficiency in Python and experience with ML frameworks (PyTorch, TensorFlow)

Recommended:

Compilers: Experience with compiler design, optimization passes, and code generation
Runtime Systems: Understanding of memory management, garbage collection, and system software
Prior exposure to EDA tools, VLSI design, or compiler optimization
Research experience in computer architecture, systems, or machine learning
Familiarity with simulation tools (gem5, Synopsys, Cadence)

Important: This course assumes you already understand how computer systems work at some level of the stack. We focus on understanding how AI systems design them, not on learning systems fundamentals. Students should have strong background in at least one area: software systems, computer architecture, or chip design.

This seminar is designed for PhD students and advanced Master’s students conducting research in computer architecture, systems, or related areas.

Course Overview

This course explores the shift from human-designed heuristics to autonomous AI agents. While conventional courses teach you how existing systems work, this seminar explores how AI agents will design tomorrow’s complete computing stack.

This course is focused on exploring how AI agents enable the specialized, efficient systems that edge computing demands. Students will systematically explore AI agent applications across the complete computing stack - from code generation and compiler optimization, through processor and accelerator design, to chip placement and verification - while identifying the most promising research directions for agent-driven computer systems design.

What Makes This Different from Traditional Architecture Courses?

Traditional Advanced Computer Architecture:

Analyze existing processor designs and memory hierarchies
Understand performance trade-offs in current systems
Study established design principles and optimization techniques

Architecture 2.0 (This Course):

Use and evaluate agentic AI systems that automatically design processors and memory hierarchies
Analyze new design methodologies using agentic approaches through hands-on assignments
Critically assess tools and frameworks for agentic architecture exploration
Synthesize research to identify future directions in agentic systems design

The Goal: By the end of this course, you won’t just understand how architectures work—you’ll have hands-on experience with agentic AI tools and deep knowledge of how they’re transforming computer systems design across the entire stack.