Writings
Methods, playbooks & field notes
How I ship, govern, and evaluate AI systems — written from products running in production, not demos.
The Most Expensive Employee Is the One That Does Nothing
Token economics in 2026 is dominated by cost-per-token talk — but for an always-on agent that watches all day and acts rarely, the wrong metric will quietly bankrupt you. The number that decides whether a monitor is viable is cost per decision, and almost nobody tracks it.
Read29 Jul 2026I Default to One Agent
Multi-agent orchestration is the premature optimization of 2026 — elegant on a whiteboard, miserable to operate. The honest position is to default to one well-contexted agent and split only on a short, specific list. Here's that list, and where one agent won in production.
Read29 Jul 2026Cognitive Debt Is the Real Cost of Agent-Written Code
The vibe-coding debate is fought over bugs and code quality — the wrong ground. The cost that compounds is cognitive debt — working code no human can fully explain. In a consumer tool that's a maintenance tax; in a regulated domain it's a liability. The fix is making the review loop a readability gate, not just a correctness one.
Read26 Jul 2026The AI Adoption Day Playbook
A facilitation runbook for turning a one-off hackathon into a measurable adoption program — cohorts, company-specific build challenges, a weighted judging rubric, and an outcomes-first retro. Written around AI-assisted development with Cursor as the reference tool; the method generalises to any agentic IDE.
Read26 Jul 2026The Evaluation Layer I Keep Rebuilding
Every agent I ship needs an answer to "is this output safe to act on?" That answer is evaluation — and I've hand-rolled it twice. Here's the pattern, and why it's the part of agent engineering that decides whether you've built a toy or a system.
Read