CMU 11-768 AI Agents：OpenHands 作者亲授，从零搭 Harness、建评测、用 RL 训 Agent，视频讲义全开放

📌 课程官网：https://www.cmu-agents.com/ YouTube 播放列表：https://www.youtube.com/playlist?list=PLSN0qpDfUvTM GitHub 组织：https://github.com/cmu-agents 授课：Graham Neubig + Daniel Fried ｜ CMU LTI ｜ 11-768 ｜ Fall 2026

---

CMU 语言技术研究所（Language Technologies Institute）2026 年秋季开了一门新课：11-768「AI Agents」。

主讲是 Graham Neubig——CMU LTI 教授，OpenHands 的核心创始人（OpenHands 就是之前的 OpenDevin，目前最活跃的开源 AI 软件工程 Agent 框架之一）。联合讲师是 Daniel Fried，CMU LTI 教授，曾在 Meta AI 做智能体研究。

这门课的特点：不是讲 Agent 是什么，是让你从零把 Agent 系统真的搭出来——Harness、评测、训练三件事全做。

上课时间：周二、周四 3:30–4:50pm ET，Porter Hall 100。

---

三阶段课程结构（28 节课）

第一阶段：Build / Evaluate / Train（Week 1–4）

Week 1（L1–L4）：Agent 总览、工具调用、上下文管理、记忆与技能

Week 2（L5–L6）：规划、代码 Agent

Week 3（L7–L10）：GUI Agent、SFT（监督微调）、深度研究 Agent

Week 4（L11–L12）：高级 RL 算法、RL 系统

第二阶段：领域扩展（Week 5–8）

Week 5（L13–L16）：沙箱机制、OpenHands 专场、LangGraph 专场、可观测性

Week 6–7（L17–L20）：工作的未来、多智能体交互、人机协作系统

Week 8（L21–L23）：树搜索、客座讲座（Karthik Narasimhan、Sasha Rush）

第三阶段：期末（Week 9）

最终 Poster 展示

---

三次作业（逐步递进）

Assignment 1（截止 9/14）：从零搭 ReAct Harness

内容：

• 实现工具调度循环
• 技能 YAML 发现机制
• 6000-token 上下文压缩
• 在真实 SWE-bench 实例上运行并评分

Starter code：https://github.com/cmu-agents/assignment-1（配套 chess-app bug 靶标应用）

这个作业的要求是：能在 SWE-bench 实例上跑通、得分。不是玩具 Demo，是工业级评测场景。

Assignment 2（截止 9/24）：设计评测框架

内容：

• LLM-as-judge 评测框架设计
• FAIL_TO_PASS 指标实现
• 对 Agent 的软件修复能力进行系统性评估

Assignment 3（截止 10/22）：训练

内容：

• SFT（监督微调）轨迹收集与训练
• RL 训练（基于奖励信号）
• 对比训练前后 Agent 的行为差异

团队研究项目：占总分 50%，选题自定，期末 Poster 展示。

---

已公开的材料

YouTube 播放列表：https://www.youtube.com/playlist?list=PLSN0qpDfUvTM

目前已上传前 4 讲，约 4.5 小时。L1 视频：https://www.youtube.com/watch?v=UwfjzyLnvMg

讲义 PDF：第 1–6 讲已公开。

GitHub：https://github.com/cmu-agents — Assignment 1 starter code（50+ stars）已开放。

官网：https://www.cmu-agents.com/ — 完整 syllabus、作业说明、讲师信息。

---

技术重点拆解

ReAct Harness ReAct（Reasoning + Acting）是目前主流 Coding Agent 的基础框架——交替进行推理和行动（工具调用）。Assignment 1 要求学生从零实现一个完整的 ReAct 循环，包括上下文压缩策略，然后在真实 SWE-bench 实例上验证。

LLM-as-judge 评测 用语言模型作为评判器评估 Agent 行为，是当前 Agent 评测的主流做法。Assignment 2 要求学生设计系统性的评测框架，不是跑跑看，而是定义可量化的指标（FAIL_TO_PASS）。

SFT + RL 训练链路 Assignment 3 覆盖了目前 Agent 训练的完整链路：先收集成功轨迹做 SFT，再用强化学习信号进一步优化。这是当前 Agent 能力提升的标准路径（类似 Devin、OpenHands 的训练方式）。

OpenHands 专场 Week 5 有一节专门讲 OpenHands 内部架构，由 Neubig 本人主讲，相当于官方拆解。

---

客座讲师

• Karthik Narasimhan — Princeton 教授，AI Agent 领域核心研究者，TextWorld 环境作者
• Sasha Rush — Cornell Tech 教授，Hugging Face 研究员，Annotated Transformer 作者

---

适合谁

• 想从工程角度系统掌握 Agent 构建的工程师
• 想了解 OpenHands 架构设计的开发者
• 需要系统理解 Agent 评测和训练的研究者

这门课的价值不在于”介绍 Agent 概念”——那类内容网上很多。价值在于：从零写、从零评、从零训，每个环节都有对应作业强制落地，而且用的是真实评测集（SWE-bench），不是玩具问题。

---

开源代码与模型仅供学习、勿直接用于工作。

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。

📌 Course website: https://www.cmu-agents.com/ YouTube playlist: https://www.youtube.com/playlist?list=PLSN0qpDfUvTM GitHub org: https://github.com/cmu-agents Instructors: Graham Neubig + Daniel Fried | CMU LTI | 11-768 | Fall 2026

---

CMU’s Language Technologies Institute opened a new course in Fall 2026: 11-768 “AI Agents.”

The primary instructor is Graham Neubig — CMU LTI professor and core co-founder of OpenHands (formerly OpenDevin, currently one of the most active open-source AI software engineering agent frameworks). The co-instructor is Daniel Fried, CMU LTI professor and former Meta AI agent researcher.

This course’s defining characteristic: it’s not about explaining what an agent is — it’s about actually building one from scratch. Harness, evaluation, and training: all three, hands-on.

---

Three-Phase Course Structure (28 Lectures)

Phase 1: Build / Evaluate / Train (Weeks 1–4)

Week 1 (L1–L4): Agent overview, tool-calling, context management, memory and skills

Week 2 (L5–L6): Planning, coding agents

Week 3 (L7–L10): GUI agents, SFT (supervised fine-tuning), deep research agents

Week 4 (L11–L12): Advanced RL algorithms, RL systems

Phase 2: Domain Extensions (Weeks 5–8)

Week 5 (L13–L16): Sandboxing, OpenHands deep dive, LangGraph session, observability

Weeks 6–7 (L17–L20): Future of work, multi-agent interaction, human-agent collaborative systems

Week 8 (L21–L23): Tree search, guest lectures (Karthik Narasimhan, Sasha Rush)

Phase 3: Finals (Week 9)

Final poster presentations

---

Three Assignments (Progressive)

Assignment 1 (due 9/14): Build a ReAct Harness from scratch

Requirements:

• Implement a tool dispatch loop
• Skill YAML discovery mechanism
• 6000-token context compression
• Run and score on real SWE-bench instances

Starter code: https://github.com/cmu-agents/assignment-1 (includes chess-app as the bug target application)

The requirement: get it working and scored on a SWE-bench instance — not a toy demo, but an industrial evaluation scenario.

Assignment 2 (due 9/24): Design an evaluation framework

Requirements:

• LLM-as-judge evaluation framework design
• FAIL_TO_PASS metric implementation
• Systematic evaluation of agent software repair capability

Assignment 3 (due 10/22): Training

Requirements:

• SFT trajectory collection and training
• RL training (reward signal-based)
• Compare agent behavior before and after training

Team research project: 50% of the final grade; topic open; final poster presentation.

---

Published Materials

YouTube playlist: https://www.youtube.com/playlist?list=PLSN0qpDfUvTM — first 4 lectures uploaded (~4.5 hours). Lecture 1: https://www.youtube.com/watch?v=UwfjzyLnvMg

Lecture PDFs: Lectures 1–6 available.

GitHub: https://github.com/cmu-agents — Assignment 1 starter code (50+ stars) open.

Official site: https://www.cmu-agents.com/ — full syllabus, assignment details, instructor information.

---

Technical Deep Dives

ReAct Harness ReAct (Reasoning + Acting) is the foundation of most current coding agents — alternating between reasoning and action (tool calls). Assignment 1 requires implementing a complete ReAct loop from scratch, including a context compression strategy, then validating on real SWE-bench instances.

LLM-as-Judge Evaluation Using a language model as a judge to evaluate agent behavior is the current mainstream approach for agent evaluation. Assignment 2 requires designing a systematic evaluation framework with quantifiable metrics (FAIL_TO_PASS) — not just “run it and see.”

SFT + RL Training Pipeline Assignment 3 covers the complete current agent training pipeline: collect successful trajectories for SFT, then optimize further with reinforcement learning signals. This is the standard path for improving agent capabilities (similar to how Devin and OpenHands are trained).

OpenHands Deep Dive Week 5 includes a dedicated session on OpenHands’ internal architecture, taught by Neubig himself — essentially an official walkthrough.

---

Guest Lecturers

• Karthik Narasimhan — Princeton professor, core researcher in AI agents, author of the TextWorld environment
• Sasha Rush — Cornell Tech professor, Hugging Face researcher, author of The Annotated Transformer

---

Who It’s For

• Engineers who want to systematically build agent systems from an engineering perspective
• Developers who want to understand OpenHands’ architectural design
• Researchers who need to systematically understand agent evaluation and training

This course’s value isn’t in “explaining agent concepts” — that content is everywhere. The value is: build from scratch, evaluate from scratch, train from scratch, with every step enforced through assignments using real evaluation sets (SWE-bench), not toy problems.

---

Open-source code and models are for learning purposes only — do not use directly in production work.

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.