Trust, Oversight and Agent Version Control

The Rise of Malleable Software 🧩 — with Geoffrey Litt

Topic 8 of 11

Published Jul 13, 2026 · Clip 33:11–40:22 (7:10)

Geoffrey explains that a simple principle—anything you can do with a person you should be able to do with an AI—goes far but breaks down around trust. Agents are less reliable than humans, creating tension between giving them large scope and injecting human review. He cites version control and GitHub as inspiration: coding agents make sense partly because PRs and reverts exist. He describes Notion's shipped feature showing exactly what changes an agent made to a page, plus reviewing agent outputs via explanation pages with diagrams and collaborative comments before reading code.