Field notes & deep-dives
Notes from building distributed systems, cloud infrastructure, and production-grade GenAI agent architectures.
Why I Built Antaeus
Turning plain-language policies into versioned, testable decisions: why Antaeus separates model judgment from deterministic application outcomes.
Context Inheritance, Published in ACM
Our AgenticDev ’26 paper, the Munich presentation, and what the study says about helping coding agents inherit repository knowledge.
Why I Built Discovery Quest
A father of three on turning inevitable screen time into a curiosity engine: a voice-first learning world guided by Luna the owl, with an AI-native open-core engine where courses are data an AI can draft and a validator can prove.
Creating Demo Videos the Agentic Way
A product demo built as code, not recorded: ElevenLabs voiceover (it can clone your own voice), Playwright screen capture seeded with real data, ffmpeg mixing with sidechain ducking, and a generated YouTube thumbnail — driven end to end by an agent, and re-rendered with one command.
Building From My Phone: Claude Code, caffeinate, and an ngrok Tunnel
Remote control drives your local machine — not the cloud. That one fact is what makes a phone-based build-and-preview loop possible: keep the Mac awake with caffeinate, tunnel the dev server with ngrok, and review on the go.
Your Context Window Is a Budget
The model is stateless, so every turn re-sends the whole conversation plus fixed overhead — a real shock to most people. Once you see context as a recurring budget, /context, /compact, /clear and /rewind stop being magic.
Parallel Claude Code: Worktrees and Named Sessions Without the Chaos
Running several agents against one repo breaks on shared state — one working tree, one branch. Here's the collision (with real git output), and the fix: a worktree per task and a name you can resume by.
The .env That Vanished: What Claude Code's deny Rules Actually Protect
I was certain my own security config was eating my .env in git worktrees — certain enough to write it up as advice. Then I instrumented the hook instead of guessing. The deny rule was a red herring; the real bug was one missing mkdir.
Local or Remote? A Decision Rule for Scheduling AI Agents
The hard question about a scheduled agent isn't what it does — it's where it runs. The deciding axis is the secret, not the task — plus why an agent's instructions belong in version control.