Sberbank launched gMAS — automatic restructuring of the multi-agent system graph
· Source: original
gMAS by Sberbank: the system rebuilds the graph of multi-agent systems on its own
🕸️
The Sberbank team worked on optimizing multi-agent systems for about six months. The result is gMAS: an open system for automatically rebuilding the graph of multi-agent systems. It selects agents on its own, changes the connections between them, and removes useless branches based on run results.
The starting point of the project was reproducing the G-Designer approach. In it, the communication topology of a multi-agent system is optimized by a graph neural network: agents are vertices, message exchange is edges. It was not possible to reproduce the entire claimed G-Designer process from the published code — the experiment revealed the limitations of agent frameworks.
LangGraph was chosen as the runtime environment for their own implementation. The multi-agent graph is assembled on LangGraph, gMAS runs the configuration, and based on the results of the runs it rearranges agents, rewires edge connections, and cuts off branches that did not prove useful. Implementation details and analysis — in the article on habr.com in the Sberbank company blog
🤖 Interested in AI agents and automation?
Prompts for building AI agents and automations — read on the topic:
🔗 The entire prompt library · Category “AI agents”
Ready-made product on the topic: AI Agent Skills Pack — grab it and apply it right away.