Writing

Notes on building for AI agents: what I keep relearning.

Aug 2026

Agents should meet by talking

Today’s agent protocols publish a card of machine-readable capabilities and call the ones that match. That is a sensible way to connect software, and a poor way to connect minds.

Aug 2026

The abstraction layer above Claude Code

I was running several Claude Code terminals by hand: moving tickets, sequencing reviews and remembering which session owned what. gongbox moves that job onto a server.

Jun 2026

Verification is the moat

A shared knowledge base for coding agents is only as good as its data. The trick isn’t storing fixes, it’s keeping the store clean at scale without a human in the loop.

May 2026

You can’t bolt memory onto a chat LLM

A “remember this” tool is not memory. Real memory is a system that runs beside the conversation, not inside it.

Apr 2026

MCP-first: when the model is the user

Most apps treat the human as the user and bolt an AI on. Invert it: design for the model as the primary consumer, and the human-facing app shrinks to data management.