Insights / Notes
Practical perspectives on AI-driven development, new ventures, development organizations, and cloud infrastructure — from both technology and business angles.
Building Your First Harness
A harness is a thin layer of scripts sitting between you and your business systems — not a robot, but a bridle. Its behavior reduces to four things: collect, reconcile, generate, monitor. Collection stays deterministic; judgment goes to AI or a human. Drawing on 40+ harnesses across two live operations, here is how to pick your first one and finish it in a day.
Read more →Don't Make "Automate Everything" the Goal
The first question in any automation conversation is "how far can this go?" — yet the harnesses that survive in production are not optimizing for automation coverage. They sort each target system into one of three connection routes, and what cannot be reached is left explicitly manual. Here are the design principles behind harnesses I actually run, and the rails that keep irreversible operations from happening by accident.
Read more →People Arrived Before the Playbook Did
I'm standing up operations for a newly built regional data center from Tokyo, on top of another role. The intended order is build the site, build the operating playbook, then stand up the organization — but in reality people arrived first and operations began with no playbook at all. Here is how 12 harnesses and 6,000 lines of AI-driven code held that together, and what they could not hold.
Read more →If the Same Fact Lives in Two Places, It Will Diverge
After multiplying business harnesses with AI assistance, what determined their effectiveness was not the quality of the scripts but deciding, in advance, where information lives. Fix what each location holds as canonical and what it must not hold, and split documentation into spec, procedure, and rationale. Here is the source-of-truth design I actually run, and the effect of treating the record itself as a deliverable.
Read more →Buy the SaaS, or Build It with AI? — The Build vs Buy Inversion
Instead of subscribing to a cloud CTI SaaS, I built one for our sales team — calls, recording, transcription, AI summaries, auto-dialing — on Twilio and edge infrastructure. Now that AI-driven development has broken the cost structure on the Build side, where has the break-even point moved? A first-person accounting, including the true cost of building.
Read more →How Much of a Business Unit Head's Job Can Move to AI?
I run a business unit of dozens of people while also shipping software myself. That only works because I moved most of my management workload onto a suite of Python automation harnesses — from attendance collection to board-report generation and HR evaluation entry. A first-person record of what moved to AI, and what stayed.
Read more →Running 20+ Apps in Production, Solo — Field Data from AI-Driven Development
Over roughly five months starting April 2026, I built and now operate more than 20 applications in production, alone. Here are the three conditions that made a one-launch-per-week pace possible, and what a shifted ceiling on individual productivity means for how development organizations should be designed.
Read more →Harness 7 Factions Report
Spring 2026 saw the sudden rise of 'harness' in AI agent circles. Before we knew it, distinct tribes had formed, revealing 7 factions. Here's an observation report on their respective claims and how they impact business and development decisions.
Read more →Understanding Harness Theory Through Minecraft
LLM performance isn't determined by the model alone, but by the harness that supports observation, memory, retry, and skill reuse. Using Minecraft as an example, we explain the essence of AI agent design in an accessible way.
Read more →AI-Driven Development: What Really Got Faster Was 'Speed to Deployment'
AI-driven development using Claude and ChatGPT has dramatically accelerated system development itself. However, what truly changed was the cycle from PoC to deployment, testing, operations, and business decision-making. Examining this from a business leader's perspective using a new video LMS system as an example.
Read more →Reconsidering Framework Selection in the LLM Era
Non-engineer executives' Vibe Coding is powerful. However, I've also faced the problem where evolution speed drops the moment we wrap it for production. Organizing thoughts on framework selection in the LLM era based on real-world experience.
Read more →