11-768: AI Agents (CMU, Fall 2026)
by Carnegie Mellon University (Graham Neubig, Daniel Fried)
Build an agent harness from scratch, evaluate it rigorously, then train the model behind it with SFT and RL.
Overview
11-768 AI Agents is a graduate course at Carnegie Mellon's Language Technologies Institute, taught by Graham Neubig and Daniel Fried in Fall 2026 (August 25 to December 3), on how LLM-based agents are built, evaluated and trained. Its 28 sessions move through five blocks: agent capabilities (tool use, context management, skills and memory, planning); application domains such as coding, GUI and research agents, together with supervised fine-tuning; reinforcement learning fundamentals and advanced techniques; safety, sandboxing and frameworks including OpenHands and LangGraph; and human-agent interaction, search methods and guest lectures from Karthik Narasimhan, Sasha Rush and Eric Wallace. Three assignments mirror that arc. Assignment 1, whose starter code is public on GitHub, has students implement a ReAct agent harness in Python: prompt construction, the tool-call loop with error recovery and step limits, tool dispatch into a Modal sandbox, skill discovery with progressive disclosure, and model-generated context compaction. They then use it to fix a bug in a chess app, submit a SWE-bench Django patch and play a deterministic chess bot, graded on a 100-point rubric with DeepSeek-V4-Flash as the default model behind an OpenAI-compatible API. Assignment 2 designs evaluations using LLM-as-judge and FAIL_TO_PASS metrics, and Assignment 3 implements supervised fine-tuning and RL pipelines; those two repositories are enrollment-only. A team research project is worth half the grade. As of early September 2026 the first four lecture recordings were on YouTube and six slide decks were public as PDFs, with more released as the term progresses.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- Fall 2026 semester (Aug 25 to Dec 3), 28 sessions of ~60-75 minutes; recordings posted as the term runs
- Provider
- Carnegie Mellon University (Graham Neubig, Daniel Fried)
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Implement a ReAct agent loop with tool dispatch, error recovery and step limits
- ✓Add skill discovery with progressive disclosure through an invoke_skill tool
- ✓Compact long agent context with model-generated summaries while preserving the original objective
- ✓Run agent tools inside Modal sandboxes and analyse the resulting token usage
- ✓Design agent evaluations using LLM-as-judge and FAIL_TO_PASS test metrics
- ✓Train agentic LLMs with supervised fine-tuning and reinforcement learning pipelines
- ✓Weigh harness engineering against model training when improving an agent's performance
Highlights
- •Bridges harness engineering, evaluation design and agent RL training in a single syllabus, topics usually taught separately
- •The public Assignment 1 ships real starter code, a SWE-bench task and a detailed 100-point rubric
- •Lectures are recorded and posted to YouTube during the term, alongside PDF slides
- •Guest lectures come from research groups (Karthik Narasimhan, Sasha Rush, Eric Wallace) rather than vendors
- •Caveat: Assignments 2 and 3 and sponsored compute credits are available only to enrolled CMU students
Who It’s For
Best For
- ✓ML engineers who train or post-train models for agentic tasks
- ✓Agent framework and harness developers who want research grounding
- ✓Graduate students and researchers entering agent RL and evaluation
- ✓Engineers deciding whether to fix an agent in the harness or in the model
Prerequisites
- •Prior experience training neural language models, the course's stated prerequisite
- •Graduate NLP coursework such as CMU 11-667 or 11-711 is recommended
- •Python proficiency, plus a Modal account and OpenAI-compatible API access for Assignment 1
FAQ
What is 11-768: AI Agents (CMU, Fall 2026)?
A graduate-level Carnegie Mellon course, taught in Fall 2026 by Graham Neubig and Daniel Fried, on the full supply side of AI agents. It is for engineers and researchers who want to build an agent harness from scratch, design agent evaluations, and train agentic LLMs with supervised fine-tuning and reinforcement learning.
Is 11-768: AI Agents (CMU, Fall 2026) free?
11-768: AI Agents (CMU, Fall 2026) is free to access.
What level is 11-768: AI Agents (CMU, Fall 2026) for?
11-768: AI Agents (CMU, Fall 2026) is aimed at a advanced audience. Recommended background: Prior experience training neural language models, the course's stated prerequisite, Graduate NLP coursework such as CMU 11-667 or 11-711 is recommended, Python proficiency, plus a Modal account and OpenAI-compatible API access for Assignment 1.
How long does 11-768: AI Agents (CMU, Fall 2026) take?
Expect roughly Fall 2026 semester (Aug 25 to Dec 3), 28 sessions of ~60-75 minutes; recordings posted as the term runs. Most learners work through it at their own pace.
What will I learn from 11-768: AI Agents (CMU, Fall 2026)?
You'll learn: Implement a ReAct agent loop with tool dispatch, error recovery and step limits; Add skill discovery with progressive disclosure through an invoke_skill tool; Compact long agent context with model-generated summaries while preserving the original objective; Run agent tools inside Modal sandboxes and analyse the resulting token usage; Design agent evaluations using LLM-as-judge and FAIL_TO_PASS test metrics; Train agentic LLMs with supervised fine-tuning and reinforcement learning pipelines; Weigh harness engineering against model training when improving an agent's performance.
Topics
Sources
This page was written from 5 sources, 4 on domains other than cmu-agents.com.