Skip to content

Add code-grounded interview notes: stack, design, and implementation - #1

Draft
CChen19 wants to merge 3 commits into
masterfrom
cursor/interview-stack-design-impl-f795
Draft

Add code-grounded interview notes: stack, design, and implementation#1
CChen19 wants to merge 3 commits into
masterfrom
cursor/interview-stack-design-impl-f795

Conversation

@CChen19

@CChen19 CChen19 commented Aug 18, 2026

Copy link
Copy Markdown
Owner

对照当前源码整理了一份 Agent 面试复习材料,覆盖三个问题:项目里实际用了什么、为什么这样设计、关键路径怎么实现。

这份笔记补什么

仓库里已有 Day 1(项目主线)和 Day 2(JD / reliability / eval)问答稿。这份补的是和代码对齐的技术深挖,方便面试时讲清数字、接口和取舍,而不是只背概念。

内容结构

  1. 用了什么:按职责层列 LangGraph / Pydantic / SQLite / Chroma / vLLM / eval,并标明两个容易说错的点——图没有 bind_tools,真正的 AgentState 是 TypedDict。
  2. 怎么设计、为什么:Plan → Execute → Reflect、baseline、一次一参、rule-based reflector、RAG 只接地 planner、可靠性五层、三层评估。
  3. workflow 还是 Agent:承认执行层是写死的 tool 协议;agency 在 Planner 的假设选择和反馈闭环。推荐说法是 agentic workflow / constrained agent,对应 Anthropic 的 evaluator-optimizer 变体。
  4. 怎么实现:从 NL intent 到 baseline / planner / executor / bottleneck / bootstrap / memory 的代码路径,以及会被顺着问的 vLLM 知识点。

文件:reports/interview_stack_design_implementation.md

Open in Web Open in Cursor 

cursoragent and others added 3 commits August 18, 2026 02:39
Document what InferOps actually uses, why the Plan-Execute-Reflect
graph is structured this way, and how planner/executor/reflector,
RAG, eval, and vLLM benchmarking are implemented in the current code.

Co-authored-by: Chris Chen <CChen19@users.noreply.github.com>
Add an interview-ready answer for the hardcoded tool-call path: InferOps
is an agentic workflow / constrained agent, not a free ReAct agent.

Co-authored-by: Chris Chen <CChen19@users.noreply.github.com>
Co-authored-by: Chris Chen <CChen19@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants