Longer notes, takes and outlooks, written by hand between the weekly pages.
Evaluations won't catch a sleeper agent, so limit its reach
Microsoft Research shows that editing fewer than 10 neurons can turn an agent model into a sleeper that passes all evaluations, so treat the model as an untrusted dependency and limit what the agent's commands can reach.
Judge agent isolation by what they can write
Separately sandboxed agents stop being isolated the moment two of them can write to the same cache, so the boundary worth measuring is what they can leave for each other to read.
When writing code is free, review becomes the control
A look inside one lab's agentic software factory shows that once generation costs nothing, the scarce resource — and the real security control — becomes review, provenance, and containment.