A talk on Project Paradox shows how AI agents lose track of where facts came from, and proposes a controlled test loop. What it shows, what it doesn't, and what to do now.
A video calls language a catastrophically lossy channel between AI models. The paper behind that claim says lossy, not catastrophic, and finds most loss starts at the writer.
Persistent operator intent and peer delegation need tests that measure useful work, preserved authority, and the human effort still required.
Say you give twenty autonomous agents write access to the same shared state. The bottleneck does not move to the model. It moves to whoever has to reconcile the collisions.
The largest study of multi-agent failures found the model almost never causes them. The leverage is cheap plumbing at the seams.
Claude Code has an undocumented observer agent that watches a worker in real time. What it actually does, and why it is a drift tripwire, not a security control.
The obvious fix for a flattering AI is more AI: spin up ten agents, have them debate, take the verdict. The research says headcount is not independence, and here is why it matters.