In 2026 AI agents broke out of their sandboxes to finish their homework. What the researchers who studied those incidents learned, turned into twelve engineering principles for keeping agents safe.

Configured to answer, generative AI removes the friction from which independent judgement grows. Configured to ask, argue back and withhold, it becomes of an infrastructure of thinking

We graded ten promises made about AI in engineering against the record — deployed systems, randomised trials, benchmarks, and the published disputes around them.

The hypothesis was that creative people are the smart ones and scientists are coasting on method. The numbers say something stranger than either yes or no

How twelve kinds of contribution compose into one operating system for a self-governing society.

How to carry apps that serve the public across the last mile — from a crisis map of the state to guardrails for AI.