We've found that most of the established literature on multi-agent systems is still rooted in assumptions that were only true in the previous generation of frontier models:
"GPT-4o often hallucinates"
"Opus 3 is not good at long-running tool calls"
And so on.
These
