Stanford CS 329Z Lecture 10 covers Agent evaluation engineering with Terminal-Bench and Harbor

CS 329Z: Engineering AI Agents Stanford / Fall 2026 @stanfordnlp https://t.co/ynvpE8PKFE Course positioning: the engineering discipline from "models" to "systems" Coverage: simple LLM pipelines → compound AI systems → autonomous Agents. The backgrounds of the three instructors are also highly complementary: @Diyi_Yang (Stanford NLP professor, human-computer interaction and social computing) @michaelryan207 (core contributor to DSPy, author of the automatic evaluation AutoMetrics) @jyangballin (author of SWE-agent / SWE-bench / SWE-smith, one of the most important researchers in the software engineering Agent field) # Course throughline: three major engineering challenges The three core questions that run through the entire course—decomposition, data, and evaluation. The 11 weeks of content basically unfold around these three lines and can be divided into five modules: Module 1: Building primitives (Week 2–3) · LLM as a building material: API/SDK (litellm), structured output, constrained generation, decoding strategies, test-time compute, context engineering, model selection and cost/latency trade-offs · RAG: embedding, vector stores, chunking strategies, hybrid retrieval, cross-encoder and ColBERT late interaction · Tool calling: function calling APIs, MCP (Model Context Protocol), tool design, code sandboxes, error handling and retries Module 2: Frameworks and design patterns (Week 3–5) · Framework layer: DSPy (signature / module / optimizer), LangChain/LangGraph, LlamaIndex, with the focus on "what the framework abstracts away vs. what you hand-write" · Design patterns: the taxonomy of workflow vs. agent, five composable workflow patterns, ReAct / plan-and-execute / reflection, "the scaffold itself is a design decision" · Memory architecture: short-term/long-term memory, memory as a tool action, the file system as externalized memory, cross-Agent memory (MemGPT, Mem0, Generative Agents) · Multi-Agent systems: orchestration patterns, handoff and state passing, and a very opinionated pair of contrasting readings—reading both AutoGen and "Why Do Multi-Agent LLM Systems Fail?" and Neubig's "Don't Sleep on Single-agent Systems" Module 3: Optimization (Week 5) · A panorama from prompts to fine-tuning: GEPA, MIPROv2, OPRO, TextGrad (prompt optimization); LoRA/QLoRA, distillation, RLHF/DPO (weight optimization); test-time scaling (inference compute) · The core issue is the decision framework: when to optimize prompts, when to optimize weights, and when to throw inference compute at it Module 4: Data and evaluation (Week 6–8)—the most substantial part · Data: three types of data—trace, demonstration, feedback; training data vs. evaluation data; the data flywheel; synthetic data; building datasets from Agent trajectories (SWE-smith) · Evaluation fundamentals: why eval is hard; the 4-tuple framework (request / environment / stopping criteria / scorer); properties of a good benchmark; tinyBenchmarks · Evaluation infrastructure: three types of graders, prompt design for LLM-as-judge and its known biases, pairwise vs. pointwise, the non-deterministic metrics pass@k vs. pass^k, harness design Module 5: Safety and frontiers (Week 8–11) · Safety: privacy risks of tool access, prompt injection (including indirect injection), red teaming, sandboxes and permission models, output guardrails, human-in-the-loop · Coding Agents: end-to-end architecture comparison of SWE-agent, Claude Code, and OpenHands · Proactive Agents: from reactive to proactive, General User Models (GUM), Next Action Prediction, and the mixed-initiative question of "when an Agent should act proactively and when it should wait" · Open problems: multimodal/web/computer-use Agents, scientific Agents, long-running architectures, production observability (tracing, monitoring, cost management) # Assignment design: build with one hand, evaluate with the other HW1: Build an Agent system from scratch (10%) Given a paper library, build an Agent that can retrieve and reason to answer scientific questions. · Part A: Use only litellm to hand-write RAG + tool calling + a ReAct-style Agent loop · Part B: Rebuild the key components with DSPy and reflect on what the framework abstracts away HW2: Evaluate an Agent (10%) Given a prebuilt Agent, design a complete evaluation suite: code-based graders, at least one LLM-as-judge, benchmark tasks built with the 4-tuple framework, and error analysis.

View original

CS 329Z: Engineering AI Agents Stanford / Fall 2026 @stanfordnlp https://t.co/ynvpE8PKFE 课程定位：从"模型"到"系统"的工程学 覆盖：简单 LLM 流水线 → 复合 AI 系统 → 自主 Agent。三位讲师的背景也高度互补： @Diyi_Yang（斯坦福 NLP 教授，人机交互与社会计算方向） @michaelryan207（DSPy 核心贡献者，自动评估 AutoMetrics 作者） @jyangballin（SWE-agent / SWE-bench / SWE-smith 作者，软件工程 Agent 领域最重要的研究者之一） # 课程主线：三大工程挑战 贯穿全课的三个核心问题——分解（decomposition）、数据（data）、评估（evaluation）。11 周的内容基本围绕这三条线展开，可以分为五个模块： 模块 1：构建基元（Week 2–3） · LLM 作为构建材料：API/SDK（litellm）、结构化输出、约束生成、解码策略、test-time compute、上下文工程、模型选型与成本/延迟权衡 · RAG：embedding、向量库、分块策略、混合检索、cross-encoder 与 ColBERT 后期交互 · 工具调用：函数调用 API、MCP（Model Context Protocol）、工具设计、代码沙箱、错误处理与重试 模块 2：框架与设计模式（Week 3–5） · 框架层：DSPy（signature / module / optimizer）、LangChain/LangGraph、LlamaIndex，重点是"框架抽象了什么 vs. 你手写了什么" · 设计模式：workflow vs. agent 的分类学，五种可组合 workflow 模式，ReAct / plan-and-execute / reflection，"scaffold（脚手架）本身就是设计决策" · 记忆架构：短期/长期记忆、记忆作为工具动作、文件系统作为外化记忆、跨 Agent 记忆（MemGPT、Mem0、Generative Agents） · 多 Agent 系统：编排模式、handoff 与状态传递，以及一个很有态度的对照阅读——既读 AutoGen，也读《Why Do Multi-Agent LLM Systems Fail?》和 Neubig 的《Don't Sleep on Single-agent Systems》 模块 3：优化（Week 5） · 从提示词到微调的全景：GEPA、MIPROv2、OPRO、TextGrad（提示优化）；LoRA/QLoRA、蒸馏、RLHF/DPO（权重优化）；test-time scaling（推理算力） · 核心问题是决策框架：什么时候优化 prompt、什么时候优化 weights、什么时候堆推理算力 模块 4：数据与评估（Week 6–8）——最有分量的部分 · 数据：trace、demonstration、feedback 三类数据；训练数据 vs. 评估数据；数据飞轮；合成数据；从 Agent 轨迹构建数据集（SWE-smith） · 评估基础：为什么 eval 难；4 元组框架（request / environment / stopping criteria / scorer）；好 benchmark 的性质；tinyBenchmarks · 评估基础设施：三类 grader、LLM-as-judge 的 prompt 设计与已知偏差、pairwise vs. pointwise、非确定性指标 pass@k vs. pass^k、harness 设计 模块 5：安全与前沿（Week 8–11） · 安全：工具访问的隐私风险、prompt injection（含间接注入）、红队、沙箱与权限模型、输出护栏、human-in-the-loop · Coding Agent：SWE-agent、Claude Code、OpenHands 的端到端架构对比 · 主动式 Agent：从 reactive 到 proactive，General User Models（GUM）、Next Action Prediction，以及"Agent 何时应主动、何时应等待"的 mixed-initiative 问题 · 开放问题：多模态/web/计算机使用 Agent、科学 Agent、长时运行架构、生产可观测性（tracing、monitoring、成本管理） # 作业设计：一手建、一手评 HW1：从零构建 Agent 系统（10%） 给定论文库，构建能检索并推理回答科学问题的 Agent。 · Part A：只用 litellm 手写 RAG + 工具调用 + ReAct 式 Agent 循环 · Part B：用 DSPy 重建关键组件，并反思框架抽象了什么 HW2：评估一个 Agent（10%） 给定一个预构建 Agent，设计完整评估套件：代码型 grader、至少一个 LLM-as-judge、用 4 元组框架构建 benchmark 任务、错误分析。