Princeton University COS 597C

COS 597C Fall 2026

AI
Agents

Princeton University

Instructor
Karthik Narasimhan
TA
Seth Karten
Meeting
Fridays, 1:20–4:10 pm
Location
Friend Center 008
Fig. 1
Task-driven agent and environment loop with evaluation A TASK enters an AGENT composed of a MODEL and HARNESS. The agent exchanges actions and observations with an external ENVIRONMENT. The resulting output flows to an EVALUATOR and then to SCORE. TASK AGENT HARNESS MODEL ENVIRONMENT INTERACTION LOOP RESULT EVALUATOR SCORE
This graduate seminar provides a hands-on exploration of AI agents: autonomous systems that perceive, reason, plan, and act in complex environments using foundation models as their cognitive backbone. The course examines the full stack of agent design, from the underlying foundation models that power agents, to scaffolding architectures that structure their behavior. Topics are a subset of: agent architectures and design patterns; planning and reasoning with LLMs; tool use and API integration; multi-agent systems and collaboration; evaluation methodologies and benchmarks for agents; safety, alignment, and robustness of agentic systems.

01

Schedule &
materials

  1. W01

    Agent architectures

    Delivered

    Readings: 1 main · 1 supplemental
  2. W02

    How an agent works—and why you cannot trust the number it produces

    Delivered

    Readings: 3 main · 3 supplemental
  3. W03

    Scoring

    Readings: 3 main · 1 supplemental
  4. W04

    How many runs is a result?

    Readings: 3 main · 2 supplemental
  5. W05

    Adversarial evaluation

    Readings: 3 main · 1 supplemental

    Supplemental

  6. W06

    Interactive and long-horizon evaluation

    Readings: 4 main · 3 supplemental
  7. W07

    Project 1 showcase

  8. Recess

    No class

    Fall recess

  9. W08

    Model + harness

    Readings: 3 main · 4 supplemental
  10. W09

    Context

    Readings: 3 main · 2 supplemental

    Supplemental

    • MemGPT
    • Context Engineering: A Survey of 1,400 Papers on Effective Context Management for LLM Agents
  11. W10

    Self-improvement

    Readings: 3 main · 2 supplemental

    Supplemental

  12. W11

    Delegation

    Readings: 3 main · 2 supplemental

    Supplemental

  13. Recess

    No class

    Thanksgiving recess

  14. W12

    Project 2 showcase

02

Grading

60

Two projects

30 points each, one in each half of the semester.

20

Hypothesis comments

Comments should be detailed and thoughtful.

10

Paper presentation

Each student presents once during the semester.

10

Participation

Attendance, participation, and discussion.

Open the private COS 597C Hypothesis group from Slack. Before annotating, confirm the sidebar shows COS 597C, not Public.

Bonus: Up to 5 additional points for active Slack participation.

03

Course projects

The course has two team projects. Teams have two or three students and may change between projects.

Weeks 1–7

Project 1

Runs through the first half of the semester and concludes before fall break.

Weeks 8+

Project 2

Begins after fall break and concludes with the W12 showcase.

Teams and themes

Use the first two classes to form teams. Project themes and details will be provided on Slack. Use Slack to find teammates, and contact the course staff if you need help joining a team.

04

Sponsors

This course is supported by partners who provide compute, tooling, and infrastructure for student projects. We are grateful for their support.