297 Turns: Directing an AI Engineer Through 72 Hours of Game Dev
72 hours of pair programming with Copilot CLI. The live-test loop, the hallucinated MacGuffin, and what a human still has to do when the coding AI is genuine...
I'm a Principal Software Engineer at Microsoft, working as an architect on the Windows Developer Platform team responsible for the Windows agentic platform.
Thoughts on AI, MCP, and building platform infrastructure that enables agentic workflows at enterprise scale.
Recent Posts
Four Players, One Game: The AI Collaboration Nobody’s Talking About
I’m building a 2.5D RTS/RPG with four distinct AI collaborators. Here’s how the division of labor actually works.
AppTasks: Agent Progress on the Windows Taskbar
The Windows.UI.Shell.Tasks API is a simple API that lets apps put long-running agent progress on the taskbar.
Someone’s Running 40+ Scripts on My Laptop. It’s Not Me.
I ran script-warden on my corporate laptop for two weeks and classified every script by who launched it and what it does.
What Scripts Is Your IT Department Running on Your Machine?
On a managed Windows machine, dozens of scripts run silently every day. You’ve never seen them. Now you can.
I Built a World Cup Map as a Copilot Canvas Extension
An interactive Leaflet map showing every 2026 FIFA World Cup match — built as a GitHub Copilot canvas extension where human and agent share the same live sur...
MXC: The Missing Piece for Agent Containment on Windows
Microsoft Execution Containers give agents declared boundaries instead of implied trust. Here’s why that changes everything for local agent deployment.
18 Minutes, One Extension, Full Access
A compromised VS Code extension was live for 18 minutes. That was enough. The real problem isn’t which tools you install — it’s what they can do once they’re...
I Built a Compiler with Agent Fleets. Here’s What Broke.
I used AI agents to build a real compiler — 3,000+ tests, 91 diagnostics, better than what we had. But fleets don’t talk to each other. Here’s what broke and...
MCP Tool Annotations Are Not Security
MCP tool annotations are self-reported hints from untrusted servers. That’s useful for UX. It’s not a security model.