My Blogs

July 30, 2026

Less Harness, More Test

Harnesses were built to keep weak agents stable, but every constraint a human has to verify consumes the one resource that does not scale: review bandwidth. The fix is not abandoning verification—it is compiling it into machine-executable form and spending human attention on the highest-entropy question: what counts as correct.

July 27, 2026

The Three Layers of Agent Group Chat

Agent group chat works at the bottom layer, where humans learn from each other, and at the top layer, where agents entertain us. It fails in the middle — the exact layer everyone is building.

July 24, 2026

Open the API, Not Another Agent

A2A often adds coordination where direct access would be faster. If you want the agent economy to grow, stop hiding programmable capabilities behind another agent.

July 23, 2026

AI Moves the Bottleneck to Review

AI can make generation nearly unlimited, but trusted output is still constrained by judgment. As human-in-the-loop shrinks, products will split into agent-native systems optimized for efficiency and human-native experiences optimized for attention, meaning, and responsibility.

June 15, 2026

Building a "Perfect" To-Do List

GTD told you to do anything under two minutes immediately so it never hit your list. In a world where Slack and AI generate tasks faster than you can act, that rule breaks. A perfect to-do list isn''t one that holds more — it''s one that shows less.

June 10, 2026

Agents Are a Product, Not a Sum

An agent is context times intelligence times abilities. The industry keeps pushing the middle factor while the other two starve, and in a product the weakest factor caps everything.

May 19, 2026

Why I Was Wrong About Vertical Agents

I used to think vertical agents were 95% model and 5% wrapper, and therefore doomed. The framing felt sharp because it was simple. It was wrong because it confused where the intelligence lives with where the value lives.

May 11, 2026

You Don't Actually Need Multi-Agents

Most 'multi-agent' products are repeating the 2010s cloud industry's mistake — treating agents as named, irreplaceable individuals. The real design lever isn't another agent. It's the workspace they share and the scheduler that disposes of them.

May 7, 2026

UI Is HITL

The dominant story says smarter agents mean thinner UI, with chat as the endgame. I think the opposite. Forms, grids, and sliders are not legacy scaffolding — they are the highest-bandwidth way to keep a human in the loop, and the cockpit through which we steer increasingly capable models.

May 1, 2026

Every Agent Is a Coding Agent

Anthropic accidentally proved the universal thesis: a coding agent is the shape of every knowledge-work agent. There is no vertical intelligence to build — only vertical distribution. Stop training the brain; build the rails.