Magnus McCune
Thoughts on software, technology, and building products.
★Featured
Multi-Agent OpenClaw on a DGX Spark: Local AI That Actually Keeps Up
Running multiple AI agents on 128GB of unified Blackwell memory with MXFP4 quantization, shared Ollama inference, and full observability.
Industrial AI Needs Better Jigs, Not Smarter Models
Why scaling AI on the shop floor isn't a model problem. It's a harness problem. Lessons from woodworking, dollhouses, and 260 tiny holes.
Multi-Agent OpenClaw on a DGX Spark: Local AI That Actually Keeps Up
Running multiple AI agents on 128GB of unified Blackwell memory with MXFP4 quantization, shared Ollama inference, and full observability.
Agent-First Repos: From DX to AX
When agents spend more time in your codebase than humans do, the conventions that matter change. Anti-ambiguity beats aesthetics.
Two Industry Shifts Are Colliding
AI models are commoditizing and agents are becoming first-class software users. The durable advantage is moving up the stack.
Industrial AI Needs Better Jigs, Not Smarter Models
Why scaling AI on the shop floor isn't a model problem. It's a harness problem. Lessons from woodworking, dollhouses, and 260 tiny holes.
Agent-Driven Development: How AI Tools Built This Blog
Using Claude Code with Serena MCP and Linear integration to build a modern web application through specialized AI agents.
From 706 Over-Tests to 52 Right-Sized Tests: Building a Scalable Test Infrastructure
How we fixed a 5-to-1 test-to-code ratio, established proper unit vs E2E boundaries, and built a test suite that actually helps development.
Building a Zero-JavaScript Blog with Island Architecture
How we built a high-performance blog that ships 0 KB of JavaScript by default using Next.js 15, React 19, and server-first architecture.