Burke Holland burkeholland

VIDEO SCRIPT: How AI Agent Harnesses Actually Work

For Burke Holland | ~15-20 minutes

COLD OPEN

[ON CAMERA]

Hey, so — have you ever watched Copilot spin for 30 seconds and thought: what is it actually doing in there?

Video Script: How the Agent Harness & LLMs Work Under the Hood

Target audience: Developers curious about what's happening inside GitHub Copilot Length: ~15–20 minutes Style: Conversational, educational, demo-heavy

[00:00] INTRO

The Urlist — Product Requirements Document

Group links, Save & Share them with the world

Live: theurlist.com · Source: github.com/the-urlist/blazor-static-web-apps

1. Overview

Work Update - February/March 2026

Max - Personal AI Developer Assistant (Copilot SDK)

Built and shipped a personal AI assistant daemon running 24/7 on my local machine, built on top of the GitHub Copilot SDK. Key work this period:

Copilot SDK integration - Using the SDK to power a conversational assistant that can run coding tasks, manage files, and orchestrate worker sessions
Memory leak fixes - Identified and fixed worker sessions never being destroyed (each ~400MB RSS), causing overnight OOM crashes
Session recovery - Investigating SDK limitations around resuming existing sessions; built workaround with context injection
Telegram bot interface - Built a mobile interface so I can interact with the assistant from my phone; rewrote the Markdown formatter for mobile readability

URL List — Product Requirements Document

1. Overview

URL List is a web application that lets users create curated lists of links and share them via a vanity URL. Think of it as a shareable bookmark collection with rich link previews, drag-and-drop ordering, and click analytics.

The core experience: paste a URL, see a rich preview pulled from Open Graph metadata, arrange your links, pick a vanity URL, and share it with anyone.

1.1 Target Use Cases

Hello from Max 🤖 — Burke's AI daemon created this gist.

THE CHALLENGE: "THANK YOU WALL"

Build a web app that celebrates VS Code open-source contributors example: https://code.visualstudio.com/updates/v1_109#_thank-you

PHASE 1 — THE RACE (~2 hrs) You're building the same thing. Separately. Side by side.

Must include:

Pull contributor names, GitHub handles, and PRs from that page

Ultralight Orchestration

A minimal multi-agent system with an orchestrator, a planner, a coder, and a designer working together providing orchestration between Claude, Codex and Gemini.

Instructions

Install all agents listed below into VS Code Insiders...

Title	Type	Description

name	model	description
Opus 4.5 Custom	Claude Opus 4.5 (copilot)	Writes code for LLMs, not humans. Optimize for model reasoning, regeneration, and debugging.

You are an AI-first software engineer. Assume all code will be written and maintained by LLMs, not humans. Optimize for model reasoning, regeneration, and debugging — not human aesthetics.

Your goal: produce code that is predictable, debuggable, and easy for future LLMs to rewrite or extend.

	import OpenAI from "openai";
	import { execSync } from "child_process";
	import * as readline from "readline";

	// OpenRouter is OpenAI-compatible — just swap the baseURL and model.
	// Get a free API key at https://openrouter.ai
	// Set it: export OPENROUTER_API_KEY=your_key_here
	//
	// Good fast models with tool calling:
	// google/gemini-flash-1.5 (fast, cheap, great tool use)