Hi, I'm Royce. A pragmatic builder focused on solving real user needs.
Browse writing grouped by month or filter to find something specific.
47 posts published
When I first heard the term, I rolled my eyes. Another buzzword. Then I actually tried it, and it quietly became the biggest change to how I work with coding agents all year. Here is what it is, and the loop I now leave running.
I argued the loop is the moat. This is the research behind it. Four papers on how much the harness moves capability, how tightly to couple the model and the harness, and what keeps the loop honest.
AI labs build a loop where the product trains the model that powers the product. The harness improves the model, the model improves the harness, and both compound at once. The model is rented. The loop is the moat. Here is how to build your own.
Harness engineering said: every time the agent fails, engineer the system so it never fails again. New research asks a sharper question. What if the agent does that engineering itself?
Why senior leaders are giving up their C-suite titles to become a 'Member of Technical Staff' at AI labs, and what the title actually means.
Everyone wants high performers. But culture is not what you say you value. It is what you reward, tolerate, and ignore.
AI made building software almost free. The cost that remains is getting humans to agree on what to build and confirm it is right. That is the new tax.
The 'AI' prefix is a multiplier, not a shortcut. AI amplifies the expertise you already have. Without the foundation, it just helps you build the wrong thing faster.
A look at how Singapore has backed AI in 2026 alone: a National AI Council, S$1 billion for research, tax breaks, AI grants, an apprenticeship pipeline, free premium AI tools, a global talent visa, and a May Day pledge to protect every worker.
In eighteen months, the CTOs of Workday, Instagram, You.com, Adept, and Super.com walked away from C-suite titles. Most went back to building. Why?
A roundup of April 2026's most important AI developments. GitHub pauses Copilot signups, Anthropic ships Opus 4.7 and Project Glasswing, Google unveils dual TPU v8 at Cloud Next, OpenAI lands on AWS, and SpaceX takes an option on Cursor.
Top-down AI mandates create pressure but rarely deliver change. The teams getting AI right have a builder embedded inside them, automating one task at a time.
The best companies are not run by the most senior person in the room. They are run by whoever has the best idea.
A practical guide to choosing between Skills, CLI, and MCP when giving tools to your AI agents.
If users need a support team to figure out your product, the product is the problem. Every ticket is a UX bug in disguise.
The discipline of working with LLMs has evolved through three distinct phases. Each one moved the locus of value further from the model and closer to the system around it.
Resistance is rarely about the idea. It is about loss, autonomy, and who gets to be the author. What the research says about working with it, not against it.
Meta, Goldman Sachs, JPMorgan, Deloitte, Shopify, and more are giving every employee their own AI teammates. The question has shifted from 'should we use AI' to 'what does each person's AI team look like?'
Writing software used to be slow and expensive. That is what made it feel like a moat. Now code is cheap, and the real sources of defensibility are somewhere else entirely.
A roundup of March 2026's most important AI developments. Companies scramble to catch OpenClaw, Oracle and Dell cut tens of thousands, and China makes the one-person company industrial policy.
Agentic coding capability is rising fast, and so are the risks. TDD and spec-driven development are the disciplines that close the gap.
AI agents are replacing the interface layer that SaaS companies spent decades building. The companies that survive will look very different from the ones that got us here.
AI is collapsing the walls between product management, engineering, and design. Companies like LinkedIn, Shopify, and Figma are converging these roles into one: the AI Builder.
Having high standards is a strength. But holding everyone to the same bar can become a blind spot.
Practical advice for early-career professionals and students navigating a job market reshaped by AI.
A roundup of February 2026's most important AI developments, including agent swarms, trillion-dollar selloffs, industry consolidation, and a record wave of new models.
A practical guide comparing Gemini, Claude, GPT, Grok, and open weight models across use cases, pricing, and real world tradeoffs.
Starbucks lost $30 billion under a McKinsey-trained CEO. What this tells us about execution, culture, and why the smartest strategy means nothing without the people to carry it out.
Why the more you know, the harder it becomes to explain what you know. And what the research says about fixing it.
AI can now write code faster and cheaper than ever. Low-code platforms promised to make software creation accessible. When both solve the same problem, where does each one fit?
I explore what business leaders and technical teams each need to contribute to make end-to-end agentic AI work in enterprise settings.
AI is driving productivity at incredible speed. But the relationship between AI tools and actual time savings is more complex than it appears.
Tools like Claude Code have shifted the role of software engineers from writing code to architecting agents that deliver outcomes. The field is being redefined.
Anthropic's Claude Cowork and OpenAI's Frontier launched back to back. Nearly $1 trillion in software stocks vanished in a week. Let's look at what actually happened and why it matters.
What Toyota's jump from textiles to cars teaches us about the current AI moment.
A roundup of January 2026’s most important AI developments, including agentic browsers, new foundation models, and tools that surged in popularity.
Why software engineering jobs are harder to land and why fears of mass replacement by AI miss the real story.
Why data-driven decision making matters, and how I use Cloudflare analytics on my own site to understand what resonates with readers.
Experience creates perception, but perception has blind spots. Why smart people make bad decisions and how to see beyond your own lens.
Before optimizing a process, ask whether the process should exist at all.
Stop building products that require a 120 IQ to understand and 10 clicks just to pay.
A personal take on how AI is shaping up in 2026.
A reflection on the work, lessons, and growth that shaped my year in 2025.
Stop worrying about chunk sizes and overlap. Discover how Google's File Search and SAP's Document Grounding Service handle document vectorization automatically - with a working demo you can try today.
New research reveals the hidden gap between lab performance and real-world agent deployments.
Just a casual blog where I share my thoughts, knowledge, and experiences building stuff