On multi-agent collaboration specifically

Yan does address the dream of agents that "talk to each other" to resolve conflicts, the way two engineers would hash out a merge conflict. His view is that in 2025, agents cannot reliably hold that kind of long-context, proactive back-and-forth. Human communication is efficient because it rests on real intelligence, and agents are not there yet. So running multiple agents in collaboration tends to produce fragile systems where decision-making is too dispersed and context is not shared thoroughly enough. He is optimistic about the long term, predicting that this will improve naturally as single agents get better at communicating with humans, but he does not think it is solved now.