Insights

Nine months on agent memory and almost nothing else, in three parts. Part 1: falling into vibe coding, building memory twice, and the night I decided to dig in.

One security PR passed through 209 review conversations on its way to merge. Ninety percent were accurate. Where the human’s job goes when accuracy is free.

The WSJ reports a rise in one-employee companies clearing seven figures, read from inside a site that one human and several agents run together.

Solid earnings, a 23% drop in a day. What happens to visit-funded sites when AI search reads your content for you, from someone who just applied to AdSense.

One paragraph of Tolkien, a million tokens, two hours of autonomous execution, a 3D film. Not a demo review, but what changes about the unit of delegation.

I built Monet to fix the agent that forgets yesterday’s conventions. How an agent that writes its own memory and reads back what mattered actually works.

What today’s LLM benchmarks actually measure, what they miss, and why we need behavioral benchmarks instead of more knowledge tests.

Two requests with near-identical input tokens, a 46x cost gap. How prompt caching really works, and why agents should be designed around structured state.

Opus prices fell 67% in a year. A weekend of napkin math on GPU costs, batching, and tokenizer overhead to estimate what Anthropic actually makes per token.

Coding agents re-read the same files and re-ask the same questions every day. What memory does for an agent, how it differs from docs, and when it pays off.