CMU 11-768 AI Agents — A Course Map for Agent Engineering

CMU 11-768 AI Agents — A Course Map for Agent Engineering

CMU’s 11-768 AI Agents course is useful to LearnAI less as “another course link” and more as a signal of what a serious agent-engineering curriculum now includes: scaffolds, evaluations, training, safety, interaction, and a final research project. The course site currently frames it as a Fall 2026 graduate course on LLM-based AI agents.

*Source: 11-768 AI Agents course site Graham Neubig announcement July source screenshot: IMG_2120.PNG*

What the course signals

The most important signal is the order of emphasis. Modern agent education is no longer just “prompting plus tools.” The course framing points toward four competencies:

Competency Student should be able to
Build a scaffold Put an LLM inside a controlled loop with state, tools, and interfaces
Build evaluations Measure whether the agent completed the task, not just whether the answer sounded plausible
Train or adapt Understand when SFT, RL, or other adaptation is appropriate
Secure the system Reason about sandboxing, adversarial behavior, and unsafe actions

That is a much stronger teaching model than a tool-of-the-week syllabus.

A student-friendly module map

The course announcement and site metadata point to a broad syllabus. For LearnAI, the useful adaptation is this map:

Week cluster 1: Agent capabilities
  tool use -> context management -> skills -> memory -> planning

Week cluster 2: Agent domains
  coding -> GUI/browser -> deep research

Week cluster 3: Training and systems
  SFT -> RL -> infrastructure -> evals

Week cluster 4: Safety and interaction
  sandboxing -> adversarial defense -> human-agent interaction

This sequence makes a good template for students because every cluster can end with a concrete build or critique assignment.

Assignment pattern

Assignment What it teaches Proof of learning
Build an agent harness Agent loop, tool schemas, state The agent completes a constrained task with logs
Create an agent evaluation Task definition, success metrics A benchmark where good and bad agents separate
Train or adapt an agent Data, reward, policy change Before/after behavior with evidence
Final research project Open-ended agent engineering Reproducible repo, report, and failure analysis

The evaluation assignment is the load-bearing one. Without it, students can ship impressive demos that fail silently.

How LearnAI should use this

Do not copy the course wholesale. Use it as a benchmark for what our own student-facing wiki should teach.

LearnAI need Adaptation
Beginner pathway Pair with AI Agent Primer and the AI Agent Book
Applied course project Require a harness, an eval, and an agent trace
Research methods Ask students to critique agent papers through the evaluation/safety lens
Faculty workshop Use the four competency blocks as a 90-minute overview

Important things to know

  • The course is tied to Fall 2026, so details may change as the semester runs.
  • Use the live course site for schedule, assignments, and posted materials before assigning anything.
  • Public materials can change during the semester. Check the live site for slides, recordings, assignments, and due dates before building a class around them.
  • The course is graduate-level. For undergraduates or non-CS students, start with vocabulary and harness labs before RL.

How LearnAI Team Could Use This

  • Curriculum audit — check whether our student wiki covers scaffold, eval, training, safety, and interaction.
  • Project rubric — grade agent projects on evidence, not demo polish: task definition, logs, safety boundaries, and evaluation design.
  • Reading sequence — assign this entry after the AI Agent Book overview so students see the academic course version of the same field.

Real-World Use Cases

Scenario Use
Designing an AI-agent course Use the competency blocks as syllabus guardrails
Planning student projects Require harness + eval + trace as minimum deliverables
Faculty AI literacy Explain why agent systems are engineering systems, not just prompts
Research group onboarding Map each student project to one course competency