Endgame Engineering
Notes from the Endgame engineering team — what we build and how.
-
Set up a macOS VM for agent-driven Claude Desktop testing
What I learned running automated Claude Desktop tests inside a macOS VM: pinning the client, isolating one connector, and the rules for driving it over VNC.
-
My Self-Improving Agent Improved Itself Out of a Job
A self-improving agent read my Claude Code transcripts for four months, wrote five rules, audited whether they actually changed behavior, and rewrote its own procedures when its audits failed.
-
Give the Agent Its Own Mac
To run an agentic research program that drives a native GUI app through hundreds of trials, we booted a headless macOS VM on the same Mac and drove it over VNC — so the agent never touches the researcher's screen.
-
Lost in Transcription
A UUID is easy to store and hard to reproduce. That's a problem when your agent has to write one.
-
We Measured What Prompt Caching Actually Saves Us. It Was 68%.
Everyone says prompt caching saves money. We pulled our own usage logs and got a real number — 68% off our biggest model, net of what the cache itself costs.
-
How We Swapped Next.js for Vite Without Rewriting the App
We moved a 236K-line production app off Next.js onto a Vite SPA in six weeks, zero downtime — by re-routing instead of rewriting, and leaning hard on AI agents.
-
This Is The Simplest Blog I've Ever Built
Having built a dozen-plus publishing systems during my career, I decided to make the most bare bones one I could for our engineering blog.