Post section on Mark Zhu's Blog

Kanban - A Multi-Agent Collaboration Design Pattern

labile.zhu@gmail.com (Mark Zhu) — Wed, 13 May 2026 23:12:15 +0800

This article uses the example of generating a modern, technology-themed Mr. Bean TV episode script based on current global events to demonstrate the Kanban-based multi-agent collaboration design pattern. The article uses Hermes Agent, but the same approach can also be applied to other agent frameworks or even your own custom agents.

Introduction

For complex AI tasks, the question is no longer whether multiple agents should collaborate, but how to make multi-agent collaboration work effectively.

OpenClaw Egress Traffic Control — A New Security Market for AI Agents

labile.zhu@gmail.com (Mark Zhu) — Tue, 24 Feb 2026 03:12:15 +0800

AI Agents are evolving from tools into autonomous executors, yet our security models still operate at the level of permissions and resource access. The real risk is no longer resource access — it is intent generation and path drift.

This article shares a proof-of-concept experiment using cgroup + TProxy + TLS SNI to explore a possible enhancement path for AI security. The next wave of the security market may be built around what I call Intent Security. The rise of AI Agents could mark the beginning of a new security cycle.

Intelligent Orchestrating the Home Lab and Smart Home by OpenClaw

labile.zhu@gmail.com (Mark Zhu) — Fri, 13 Feb 2026 03:12:15 +0800

This article demonstrates how OpenClaw Skills can orchestrate Home Lab and Smart Home intelligently. By encapsulating infrastructure knowledge, monitoring systems, ann automation APIs to structured Skills. An AI agent can troubleshooting servers, visualize operations by smart lights, and deliver a summarized real time web news through smart speaker.

Background

I have been running a home lab and a small home set up for years. Ask more and more devices and services were add, maintenance and management gradually become increasingly complex. Meanwhile, after years of experience, I realized that traditional smart homes are not fully “smart”. Most of the problems exist simply because I was too lazy to automate them probably. Now, modern AI may help mitigate those issues.

The Concepts Anatomy of OpenClaw

labile.zhu@gmail.com (Mark Zhu) — Mon, 09 Feb 2026 03:12:15 +0800

OpenClaw is the most populated self-hosted personal AI assistant in the world right now. From the perspective of an AI Agent developer, it is not only a AI tool, but also one of the best real-world references for agent design patterns. It helps us create more sophisticated designs when building our agents. In this article, I will go down the rabbit hole and ultimately build a conceptual model.

Conceptual Model

The following diagram is the conceptual model I built.

Self-hosted Matrix Channel for OpenClaw - Take the red pill

labile.zhu@gmail.com (Mark Zhu) — Fri, 06 Feb 2026 03:12:15 +0800

This article explains how to self-host an OpenClaw AI assistant using a Matrix chat channel. It covers setting up a Matrix server, installing and troubleshooting the OpenClaw Matrix plugin, configuring access policies, and enabling secure, private AI interactions without relying on third-party messaging platforms.

AI-Based Job Position Watching from Company Career Pages(PoC) - Part 1

labile.zhu@gmail.com (Mark Zhu) — Mon, 19 Jan 2026 15:12:15 +0900

As an unemployed mid‑career software engineer looking for a new role in 2026, I often find myself repeatedly checking the career pages of companies I am interested in. This process is time‑consuming and tedious, especially when tracking openings across multiple companies. While many career sites provide email‑based job alerts, these alerts are usually driven by opaque AI matching against a previously submitted CV, or by simple keyword matching. In both cases, I have very limited control over the actual matching criteria.

A First Look at Proxying MCP Traffic with Agentgateway

labile.zhu@gmail.com (Mark Zhu) — Sat, 10 Jan 2026 03:12:15 +0800

As a key protocol in the AI Agent ecosystem, MCP support is one of Agentgateway’s core capabilities. This article documents my experience and observations while proxying MCP traffic with Agentgateway. It is also a summary of lessons learned after spending several days debugging why my browser could not connect to the Agentgateway MCP endpoint.

The official Agentgateway documentation does describe how to configure MCP, but at the moment the content is fairly “hello world”–level. In practice, most of my exploration involved reading the Agentgateway source code and relevant standards, including specifications related to W3C and MCP. In some cases, I also had to use Chrome DevTools to debug JavaScript and network traffic.

Agentgateway Implementation Analysis Part 3 - Http Proxy

labile.zhu@gmail.com (Mark Zhu) — Thu, 25 Dec 2025 03:12:15 +0800

This article attempts to analyze the implementation details of the Http Proxy main flow in Agentgateway. It allows readers to understand the working principle and implementation method of Agentgateway as an Http Proxy at the L7 layer. Agentgateway is essentially an HTTP Proxy, but adds support for AI (LLM/MCP/A2A) stateful protocols on top of HTTP. Therefore, analyzing the main flow of the HTTP Proxy layer is analyzing the main flow of Agentgateway.

AI Agent Bus Gateway agentgateway Implementation Analysis Part 2 - Service Lifecycle

labile.zhu@gmail.com (Mark Zhu) — Tue, 09 Sep 2025 03:12:15 +0800

Overview

This article analyzes the source code of the AI Agent bus gateway agentgateway and attempts to explain the components and collaboration related to service lifecycle management. From the startup and initialization of each service component, port listening, to how service termination signals are propagated between components, and the implementation related to graceful service shutdown (Drain).

Due to my limited understanding of Rust, especially the asynchronous programming style of tokio, please point out any errors or omissions.

Analysis of the AI Agent Bus Gateway 'agentgateway' Implementation, Part 1

labile.zhu@gmail.com (Mark Zhu) — Fri, 05 Sep 2025 03:12:15 +0800

Figure: agentgateway proxies the outbound connections of an Agent, including the MCP server, AI Agent, and OpenAPI

(source: https://agentgateway.dev/docs/about/architecture/)

Overview

This article analyzes the source code of agentgateway to provide a preliminary understanding of its main initialization process. It aims to offer a high-level reference and guidance for readers interested in diving deeper into the implementation of agentgateway.

Analyzing gemini-cli’s Source Code with gemini-cli

labile.zhu@gmail.com (Mark Zhu) — Thu, 21 Aug 2025 03:12:15 +0800

Overview

This article introduces how to use the gemini-cli command-line tool to analyze its own source code repository. Through a series of interactive commands, we will explore the design and implementation of gemini-cli. This approach not only demonstrates the functionality of AI tools but also provides developers with a new way of learning and exploration.

If you don’t know what gemini-cli is, here’s a brief introduction:
gemini-cli is an open-source Command-Line Interface (CLI) tool that allows developers and users to interact directly with Google’s Gemini model in the terminal. You can think of it as an intelligent AI assistant inside the command line.

Envoy Proxy Insider - Event-Driven and Threading Design

labile.zhu@gmail.com (Mark Zhu) — Thu, 26 Jun 2025 23:12:15 +0800

This article is an excerpt from some recent updates to my book, Envoy Proxy Insider, covering the Event-Driven Framework and the Threading Model. These are considered the foundational core of Envoy Proxy. Most people think of Envoy as a proxy that primarily forwards requests with custom logic. This is correct. However, like other middleware with high-load, low-latency requirements, its design must consider load scheduling and flow control. A good scheduling design must balance throughput, response time, and resource consumption (footprint). This article primarily discusses topics related to events, scheduling, and multi-threaded coordination.

My Book - Envoy Proxy Insider

labile.zhu@gmail.com (Mark Zhu) — Fri, 13 Jun 2025 23:12:15 +0800

This article introduces my new open-source book - “Envoy Proxy Insider”. It’s a book that dives deep into the mechanisms and implementations of Envoy Proxy. Much of its content comes from my previous book “Istio & Envoy Insider”, from which I’ve extracted the Envoy part and reorganized it into a standalone book. The purpose is to help readers focus solely on Envoy without distraction from Istio. Some content is irrelevant or inapplicable to Istio, so it makes sense to split them. Moving forward, “Istio & Envoy Insider” will only update Istio-related content, while this new book will carry all updates related to Envoy.

ChatBlog - An Experiment in AI Interaction for Personal Blogs

labile.zhu@gmail.com (Mark Zhu) — Sat, 29 Mar 2025 03:12:15 +0800

Introduction

Adding a chatbot to a blog is both interesting and practical. It can answer readers’ simple questions based on the blog content and the author’s introduction articles. ChatBlog (chat + blog) might be an engaging and useful model for interactive technical sharing. This article documents my exploration and implementation process.

Currently, the interactive knowledge base + LLM model is popular. A blog serves as an open knowledge base for tech enthusiasts, and making it open to chat is an intriguing idea.

Deep dive into Istio Ambient implement cross pods hub by network namespace trick

labile.zhu@gmail.com (Mark Zhu) — Sat, 03 Aug 2024 23:12:15 +0800

The so-called Sidecarless of Istio Ambient is, strictly speaking, a change from a sidecar container of pod to a sidecar pod of pods on a worker node. Note that I’m introducing the term sidecar pod here. To implement pods on the same worker node share a sidecar pod, you need to solve the problem of redirecting traffic from all pods to the sidecar pod.

The solution to this problem has gone through two versions during the development of Istio Ambient.

Please prove that it is the network's fault - the lost Envoy sidecar

labile.zhu@gmail.com (Mark Zhu) — Wed, 12 Jun 2024 08:12:15 +0800

Queen Christina of Sweden (left) and René Descartes

There are 3 treasures for SRE / Supporting / Performance test team in distress of troubleshooting:

Network Problems
Reboot
Hardware failure

If you meet a smart partner, he/she will be pursued: “Please use the data to prove your claim”. If you say it’s a network problem, how do you prove it?

Flow control and backpressure of Envoy/Istio

labile.zhu@gmail.com (Mark Zhu) — Sun, 21 Apr 2024 23:12:15 +0800

📚 Excerpt From:
This article is excerpted from the Flow Control section in an open source book I’m writing: Istio & Envoy Insider.
If the figures in this article not clear, you can go back to the original book.
My book: Istio & Envoy Insider includes: Envoy source code deep dive, in-depth Envoy fundamentals analysis , Istio fundamentals analysis. But it’s not a traditional “deep dive xyz source code” type of book. on the contrary, I have done my best not to directly paste source code in the book. Reading source code is a necessary step to grasp the details of the implementation, but browsing source code in a book is generally a very bad experience. So, this book uses source code navigation diagrams to let readers understand the full picture of the implementation, rather than getting lost in the details of fragmented source code snippets and forgetting the whole picture.

My new book - Istio & Envoy Insider

labile.zhu@gmail.com (Mark Zhu) — Sun, 21 Apr 2024 23:12:15 +0800

My new open source book - Istio and Envoy Insider, has just reached a milestone. So I want to Introduce it.

Overview of this book

This book is called Istio & Envoy Insider. It is a book in progress, now in draft stage.

What this book is about

This book includes: Envoy source code deep dive, in-depth Envoy fundamentals analysis , Istio fundamentals analysis. But it’s not a traditional “deep dive xyz source code” type of book. on the contrary, I have done my best not to directly paste source code in the book. Reading source code is a necessary step to grasp the details of the implementation, but browsing source code in a book is generally a very bad experience. So, this book uses source code navigation diagrams to let readers understand the full picture of the implementation, rather than getting lost in the details of fragmented source code snippets and forgetting the whole picture.

Remote debugging an Envoy sidecar running in an Istio mesh by lldb at C++ source code level

labile.zhu@gmail.com (Mark Zhu) — Wed, 10 May 2023 17:12:15 +0800

Debugging Envoy sidecar C++ code in an Istio mesh

This article is source from my open source book [Istio Insider].

Introduction

Debugging Envoy sidecar C++ code running in Istio mesh. It helps deep dive into the sidecar at code level. It makes us more confident when troubleshooting Istio problem or writing better EnvoyFilter or eBPF trace program. This article guides how to use VSCode and lldb to debug Envoy istio-proxy sidecar.

Illustrates Istio & Envoy request processing flow, timeout, circuit breaker and metrics monitoring

labile.zhu@gmail.com (Mark Zhu) — Sun, 02 Oct 2022 23:12:15 +0800

💂 About the cover:
Tower Bridge watercolor painting by Juan Bosco

📚 Excerpt From:
This article is excerpted from a section Envoy Request and Response Scheduling from an open source book I’m writing Istio & Envoy Insider If the reproduced picture is not clear, you can go back to the original book.

Request and Response scheduling

Request and Response scheduling

🎤 Before the official compilation. I would like to talk about some of the reasons for writing this chapter. Why study Envoy’s request and response scheduling?

BPF tracing istio/Envoy - Part4: Upstream/Downstream Event-Driven Collaboration of Envoy@Istio

labile.zhu@gmail.com (Mark Zhu) — Thu, 07 Apr 2022 16:12:15 +0900

Notice: The original article come from https://blog.mygraphql.com/en/posts/low-tec/trace/trace-istio/trace-istio-part4/ . If picture unclear, please redirect to the original article.

There is Chinese version too.

Why

I can’t believe that I really insisted on writing Part 4. Believe it or not, the “Why” section of each part is the hardest thing to write :) . If you’re reading this series for the first time, don’t worry, each section is relatively independent.

Introduction to Envoy you have saw earlier describes envoy’s features like this:

My real-life Istio Performance Tuning Part 2 - Starvating threads and SO_REUSEPORT socket

labile.zhu@gmail.com (Mark Zhu) — Sat, 04 Sep 2021 15:12:15 +0900

From: https://getboulder.com/boulder-artist-rocks-the-world/

Overview

Notice: The original article come from https://blog.mygraphql.com/en/posts/cloud/istio/istio-tunning/istio-thread-balance/ . If picture unclear, please redirect to the original article.

There is Chinese version too.

For a long time, programmers relied on Moore’s Law. And before it hits the ground, programmers find another lifesaver: parallelism/concurrency/eventual consistency. Today, Cloud Native/Micro-Service everywhere. Multi-threading by default. In the world of computer performance engineering, there is also a word: Mechanical Sympathy. The requirement of “Sympathy” is to understand how it work. In daily life, many people understand and pursue work life balance. But your thread, is it balance? Do you want to sympathize with it? What kind of sympathy is a thread that is tired enough to be overloaded and sees other companions eating afternoon tea? How can I get multithreading to maximum throughput?

My real-life Istio Performance Tuning - Part 1

labile.zhu@gmail.com (Mark Zhu) — Fri, 02 Jul 2021 15:12:15 +0900

Why tuning

If it takes interest and momentum to push a new technology, then what is needed to get this technology go to live production is hard working. Including Cloud Native and Istio.
In pre-launch performance testing, Istio out-of-box provided observability, operational convenience, and also introduced pain: Increased service latency. How to minimize latency has become a critical task.

My Scenario:

SERVICE-A invoke SERVICE-B
Latency of Service-A, which was previously 9ms at P50, is now 14ms at P50.

Bottleneck Analysis

There are two paths under my feet: