Six months ago, we were still debating whether 'AI could code'; now, companies are racing to see whose AI can handle more tasks simultaneously. The biggest shift isn’t that the models got stronger, but that AIs have evolved from being solo players to a complete 'execution team'.
What's happening now: First, memory mechanisms are genuinely operational; AIs no longer forget previous context every time. Second, cross-framework collaboration has been standardized; regardless of the framework used, they can interconnect through protocol standards. Third, companies are actually deploying AIs in the production phase, not just POCs.
However, this brings up a problem—most companies are still using the '1+N' model, which is one main AI + several assistants. Very few have truly designed their entire business processes to be multi-agent games and collaborations. Those that do see process efficiencies improve by 30-50%.
If you're looking to build a multi-agent system, the most practical advice right now is: don’t rush to stack up the number of AIs; first, fine-tune the memory and tools of the individual AIs, then think about how to make them collaborate. Many failed projects have done the opposite.
What's happening now: First, memory mechanisms are genuinely operational; AIs no longer forget previous context every time. Second, cross-framework collaboration has been standardized; regardless of the framework used, they can interconnect through protocol standards. Third, companies are actually deploying AIs in the production phase, not just POCs.
However, this brings up a problem—most companies are still using the '1+N' model, which is one main AI + several assistants. Very few have truly designed their entire business processes to be multi-agent games and collaborations. Those that do see process efficiencies improve by 30-50%.
If you're looking to build a multi-agent system, the most practical advice right now is: don’t rush to stack up the number of AIs; first, fine-tune the memory and tools of the individual AIs, then think about how to make them collaborate. Many failed projects have done the opposite.