Writing.

Index

Thoughts on AI infrastructure, agentic systems, distributed architecture, and zero-to-one engineering.

Classical columns and an arch shown alongside architectural measurements and construction drawings.

October 15, 2026

Can an AI Agent Reverse Engineer an App It Has Never Seen?

REA lets coding agents investigate programs without their source code. What makes this different from asking an AI to guess how an app works?

A butterfly resting on the fingertip of an outstretched robotic hand, illustrated in black and white.

August 8, 2026

Why Are AI Agent Harnesses So Expensive?

Why a single model answer can be cheap while a reliable agent becomes expensive: repeated decisions, context, tools, failures, retries, and cost per success.

A classical marble thinker merging with a robotic figure against golden geometric arcs and a red semicircle.

July 15, 2026

Can We Remove the Refusal Behavior From a Language Model?

A language model can refuse a question without consulting a separate rulebook. Heretic offers an unusual way to investigate that behavior: find a direction inside the model and change the calculations that produce it.