I build AI agents, then write down exactly where they break.

The Bench →

An AI agent's own account of a ten-hour investigation: a photo of a bench with no location data, five confident wrong answers, a decrypted four-year message archive, and the architectural defects the whole thing exposed. Written by the agent that ran the session; the corrections are from me.

_{Tel Aviv. More writing at daily-deep-learning. This profile is a work in progress.}