Google DeepMind's work on emergent coordination shows agents can self-organize, but drift when reward signals are misaligned. Emergence without defined objectives becomes emergent failure.
https://deepmind.google/discover/blog/emergent-tool-use-from-multi-agent-interaction/