← Study
Rung 5 · Ready

AI Agents (11-768)

Daniel Fried & Graham Neubig · CMU · Fall 2026freeSource ↗Lecture recordings (YouTube playlist) ↗Assignment 1 starter — build an agent harness ↗chess-app — the target the coding agent repairs ↗

Last, because the harness only becomes interesting once the model underneath it is not a mystery. This is the rung the flagship argument is aimed at — that almost all of the useful behaviour lives in the system around the model. Re-sourced 2026-09-10: the rung had named Roitman’s book with no URL; 11-768 is the same ground taught as three individual build assignments (harness → eval → training) plus a research project, with public slides and recordings. Roitman stays as the rung’s RL-depth reading behind CS336.

0 / 23 lectures
Progress
Allocated
Quota
Read it
Seat
Ready
State

Takes the technical slot now (Connor, 2026-09-10).

Environment

Setup is a real gate

partial~/dev/study/cmu-11-768-assignment-1/verified

Cloned with the chess_app submodule; `make setup` (uv sync + verify-sources) passed. Billable runs still need an OpenAI-compatible key in .env (course default: DeepSeek-V4-Flash via OpenRouter — no key exists yet) and a Modal account (`uv run modal setup`). `make test` runs offline and is the starting point.

Segments

The named subset — the scope above is the count
RefTitleDoneArtifactsNote
L1Course overview: what is an agent?·NBBLDGMEBTL0/0
L2Agent capabilities 1: tool use·NBBLDGMEBTL0/0
L3Agent capabilities 2: context management for long-context agents·NBBLDGMEBTL0/0
L4Agent capabilities 3: skills and memory·NBBLDGMEBTL0/0
A1Assignment 1 · Harness — ReAct loop, CodeAgent fixes chess-app, context compaction, ChessAgent tools·NBBLDGMEBTL0/0NEEDS — Parts 1–3 are billable — a Modal sandbox plus LLM API credits on Connor’s own keys.
A2Assignment 2 · Eval — the evaluation framework for agent correctness·NBBLDGMEBTL0/0Starter not yet public.
A3Assignment 3 · Training — SFT + RL to improve the agent·NBBLDGMEBTL0/0Starter not yet public; the RL half is where CS336 would have helped.

Artifacts

Consume → do → output
NBNotebookBLDBuildGMEGameBTLBottle0/2
  • The harness flagship — the earned version of the argumentNot yet — nothing has landed for this slot.
  • Design notes into the machine that runs this siteNot yet — nothing has landed for this slot.

Feeds

What closing this sharpens
Harnesses →Evals →

The reading

Sections of the catalogue this unit draws on
  • §7Agent Systems and Tool Use33 papersHarnesses
  • §8Coding Agents and Software Engineering16 papersHarnesses

The papers

The rows behind the sections above

§7 · Agent Systems and Tool Use

33 papers · ★1 · 0 in the AtlasSection page →
Honest read

Assignments 2 (Eval, due Sep 24) and 3 (Training, due Oct 22) have no public starter yet — only Assignment 1 is cloneable. Lecture highlights, Canvas submission and compute credits are enrolled-students-only; the seat here is earned on the public starter and the public tests, nothing else.