WAIC AI Journey session spotlights next-gen agent architectures

This WAIC session reads like a roadmap for what comes after today’s monolithic transformers. Sberbank’s AI leadership framed the future as an “economy of autonomous agents,” where graph‑structured multi‑agent systems handle transactions and physical processes while humans orchestrate at a higher level. Russian and Chinese researchers argued that scaling massive pre‑training plus RLHF is running into data and cost ceilings, and pitched alternatives like asynchronous recurrent architectures (RWKV‑7) and agent‑based graphs (gMAS) that learn continuously and self‑organize. ([businessnewsthisweek.com](https://businessnewsthisweek.com/technology/waics-ai-journey-session-addresses-changing-role-of-ai/))

What’s interesting is who is saying this. Sber is a systemically important bank, not a niche lab, and Skoltech’s AI Center and China’s RWKV/YuanShi community are increasingly influential in open‑weight model circles. Their message is that “true autonomy” will come from decentralised swarms of specialized agents rather than ever‑bigger single models. That aligns with a broader industry turn toward tool‑using, planning agents and away from chatbots as the primary interface. ([businessnewsthisweek.com](https://businessnewsthisweek.com/technology/waics-ai-journey-session-addresses-changing-role-of-ai/))

If these ideas gain traction, the AGI race becomes less about one model “winning” and more about which ecosystem can coordinate huge numbers of agents safely and efficiently. That raises new research and governance problems — from credit assignment in agent swarms to monitoring emergent behaviors in gMAS graphs — that today’s benchmarks barely touch.