Recent Writing10 entries
  1. I Started with a while Loop and Built MiniCode—an Agent That Asks First

    MiniCode is a lightweight coding agent I built from Pi's source code and several real-world lessons. It can read code, propose patches, and run checks, while guided mode pauses for human confirmation before planning, writing, or executing commands.

    MiniCodeCoding AgentTypeScriptOpen Source
  2. I Didn't Understand How Coding Agents Work Until I Ran Pi

    From the agent loop and tool adaptation to extension policies, sessions, and concurrency. I went beyond reading the code and built path-protection and audit extensions myself; a missing audit result ended up revealing Pi's event boundaries with surprising clarity.

    PiCoding AgentsAI Agents
  3. The Foundations I Added to selfweb, Starting with a Background Image

    From homepage background loading and a shared system for articles and Notes to Sanity publishing and cache revalidation, this update filled in several links that are most likely to break when a personal website is meant to support long-term writing.

    Personal WebsiteNext.jsSanityPerformance Optimization
  4. Lessons from Building a PvZ Clone During My Internship

    A look at several small but easy-to-miss issues in a Pygame PvZ clone—opening timing, cone and bucket armor, pea targeting, and zombie blocking checks.

    PygameGame DevelopmentDebug
  5. From Neural Networks to Transformers

    At heart, AI chat is a relentless next-word prediction machine. Starting with handwritten-digit recognition, this article works up to vectors, attention, and temperature in Transformers.

    TransformerNeural Network
  6. Inside OpenHuman: The Memory Architecture Behind an Open-Source Personal AI

    A three-layer summary tree replaces full-corpus RAG search, with a look at its hierarchy, comparisons with Mem0 and MemGPT, and the experience behind two merged contributions.

    OpenHumanMemory
  7. 3K Lines, 9 Tools: I Read All of GenericAgent's Source Code

    Nine atomic tools, a four-layer memory architecture, and Working Memory injection: how 3K lines of code achieve full-system control with one-third of the token budget.

    GenericAgentAgent
  8. A Deep Dive into Claude Managed Agents

    A complete look at Anthropic's newly released managed Agent platform, from architecture to hands-on use. As the platform absorbs the infrastructure, the competitive focus for Agent developers is moving up the stack.

    ClaudeAgent
  9. Dify Source Walkthrough 02: Agent Runtime

    A layer-by-layer walkthrough of Dify's three-tier Agent architecture, from graphon to the workflow glue layer and the legacy Agent runner, compared with the design gaps in my own LangGraph system.

    DifyAgent
  10. Dify Source Walkthrough 01: RAG Pipeline

    After spending two days taking apart Dify's RAG Pipeline, I found that the gap between an industrial system and a practice project lies less in the algorithms than in engineering details such as parent-child chunking, concurrent retrieval, and fast failure.

    DifyRAG