Advanced Software Development with AI

Overview

This course explores how AI is reshaping software engineering practice. We'll start with LLM foundations and core agent capabilities like context, tools, and reasoning, then move into building and training real coding agents. From there, we'll turn to evaluating, verifying, and securing those agents through benchmarks, trustworthiness, and red teaming, wrapping up with final project presentations.

Course Details

• Term: Fall 2026
• Instructor: Baishakhi Ray
• Schedule: Fridays, 10:10am to 12:40pm
• Location: 451 Computer Science Building
• Course Number: COMS W4995, Section 004
• Points: 3
• Method of Instruction: In-Person

Course Links

• Vergil
• Ed Stem
• Canvas

Announcements

Instructors

Baishakhi Ray

Teaching Assistants

Ira Ceka (Head TA)

• Email: iceka@cs.columbia.edu
• Office Hours: By appointment.
  Booking calendar on Ed Discussion.

Alex Mathai

• Email: alexmathai@cs.columbia.edu
• Office Hours: Mondays, 8:30-10pm; Fridays, 6-8pm.
  Link on Ed Discussion.

Syllabus

Date	Topic
Sep 11, 2026	IntroSlides
Sep 18, 2026	Coding AgentsSlidesPaper 1Paper 2 Supplementary Materials General info and a sample split of how presenters might divide up the work.Sample Presenter Schedule Example of slides covering introduction, methods, evaluation, and conclusions.Sample Slides Example of slides covering related work, critique, and discussion.Sample Slides Presenter Materials Slides from the student presentation of Paper 1.Paper 1 Slides
Sep 25, 2026	Memory & ContextPaper 1Paper 2 Supplementary Materials Supplementary readings for Paper 1."HiAgent: Hierarchical Working Memory Management for Solving Long-Horizon Agent Tasks with Large Language Model"Reading 1"MemGPT: Towards LLMs as Operating Systems"Reading 2 Supplementary readings for Paper 2."Context as a Tool: Context Management for Long-Horizon SWE-Agents"Reading 1"Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models"Reading 2
Oct 2, 2026	Training (Part 1)Paper 1Paper 2 Supplementary Materials Supplementary readings."DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning"Reading 1"Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?"Reading 2"Training Language Models to Generate Quality Code with Program Analysis Feedback"Reading 3
Oct 9, 2026	No Class
Oct 16, 2026	Training (Part 2)Paper 1 Paper 2 SWE-benchSWE-smith
Oct 23, 2026	EvaluationPaper 1Paper 2
Oct 30, 2026	Verification & TrustworthinessPaper 1Paper 2
Nov 6, 2026	Security & SafetyPaper 1Paper 2
Nov 13, 2026	Red TeamingPaper 1Paper 2
Nov 20, 2026	Advanced TopicsPaper 1Paper 2
Nov 27, 2026	No Class (Thanksgiving Recess)
Dec 4, 2026	Final Presentations

Logistics

Date	Event
Sep 29, 2026	Milestone 0 released
Oct 6, 2026	Milestone 0 due
Oct 16, 2026	Milestone 1 released
Oct 27, 2026	Milestone 1 due

Late Policy

• Extensions must be approved by the instructor, not the TAs.
• Approved late submissions receive at least a 10% penalty.

Attendance & Presentations

• Presentations and attendance are 40% of the final grade.
• Attendance is mandatory. Arrive on time and scan the QR code in class.
  • Submit the evaluation during class, or your attendance will not be counted.
• Coordinate with your teammates and check the sign up sheet and syllabus for your paper.
• You will get an email before your presentation.
  • Reply by Monday of that week to confirm.
  • After presenting, upload your slides to the shared drive folder.

Course Outcomes

• Explain how LLMs and agents work
• Build an agent from scratch
• Critically assess research spanning recent advancements in the field
• Understand verification layers and trustworthiness in agentic code generation
• Conduct adversarial red teaming to identify vulnerabilities and harden agentic security