With multi-agent development, each agent generates code that no single human fully understands. The failure mode isn't a consensus problem or a network partition. It's a comprehension partition. Five agents each wrote part of the system. None of them hold a mental model of the whole. Neither does the human who orchestrated them. When it breaks, there is no call chain to trace because there was never a unified understanding of what the system was supposed to do at that level of specificity.
"Done, merged to develop".
I test, feature not there.
"?" Claude: "Yeah, there's nothing for that feature in develop"
"I'm confused. You said above you merged it into develop." Claude: "I did say that but I didn't do it. Should I do it now?".
Me, thinking, "That depends, will you actually do it now?"