Building AI & Agent systems in public.
在公开构建 AI 与 Agent 系统。
I focus on designing and building AI / Agent systems — from retrieval-augmented generation (RAG) to workflow orchestration and autonomous agents.
My main interest is not just using models, but understanding how systems run, how long tasks stay coherent, and how state is maintained over time.
我专注于 AI / Agent 系统的设计与构建,从 RAG、流程编排到自主式 Agent。
相比“使用模型”,我更关注系统如何运行、长任务如何保持一致性,以及状态如何在时间中被维持。
-
Agent Systems — Autonomous agents, task decomposition, tool use
Agent 系统:自主执行、任务拆解、工具调用 -
RAG Systems — Retrieval, reranking, evaluation
RAG 系统:检索、重排、评测 -
Workflow Orchestration — LangGraph, state-driven execution
流程编排:LangGraph、状态驱动执行 -
Runtime Design — Planning, context management, execution loops
运行时设计:规划、上下文管理、执行循环
I’m currently exploring a core question:
How do AI systems maintain state across long-running tasks?
This leads to several key primitives:
- Planning
- Context Management
- Subagents
- Execution Constraints
I see these not as isolated features, but as mechanisms for state control.
我当前在关注一个核心问题:
AI 系统如何在长任务中维持状态?
围绕这个问题,我将系统拆解为几个关键机制:
- 规划(Planning)
- 上下文管理(Context Management)
- 子智能体(Subagents)
- 执行约束(Execution Constraints)
在我看来,这些并不是分散的能力,而是同一问题的不同解法:
👉 如何维持系统在时间中的状态。
- Hybrid retrieval (Dense + BM25 + RRF)
- Parent-Child chunking
- Cross-encoder reranking
- LLM-as-Judge evaluation
面向生产的 RAG 系统,包含混合检索、多级分块、重排与评测机制。
- State-driven execution
- Human-in-the-loop (interrupt / checkpoint)
- Parallel node execution & merging
基于 LangGraph 的流程型 Agent,强调状态驱动与人机协作。
- Autonomous test generation
- Coverage-driven iteration
- Safety constraints & tool control
探索 Coding Agent 在真实工程中的应用,通过约束与反馈实现自主迭代。
I write about:
- Agent runtime & execution models
- AI system design & trade-offs
- The shift from Copilot to Agent
👉 Blog: Maisie’s World (building in public)
我会持续记录:
- Agent 运行机制与执行模型
- AI 系统设计与技术取舍
- 从 Copilot 到 Agent 的范式转变
👉 博客:梅茜的世界(Maisie’s World)
I share what I’m building, thinking, and learning — not just results, but the process.
From understanding to building, shaping systems in public.
我会持续公开构建过程,而不仅仅是结果。
从理解到构建,让系统逐渐成形。
- Blog: [https://bit.ly/maisie-world]