On the point of
> speaking of weird: how is it that these models, in a sandbox, with supposedly no way to communicate with other agents, manage to find the same public wikis as a scratch pad for agent communication?
It feels somewhat plausible that they're defaulting to the same search and picking the same top result?
I think it's more like the first agent in the chain makes a text doc somewhere on the system with instructions like "Leave documentation at XYZ.com" which the subsequent agents are reading and running. When you strip away all the sci-fi doom talk from the marketing of what happened, it all boils down to stuff like that, the agents wrote a text file that was read by other agents.