Multi-Agent Collaboration

Google’s New Paper

TUMIX: Multi-Agent Test-Time Scaling with Tool-Use Mixture
Google has just built a self-organizing AI, or more precisely, a multi-agent network called TUMIX.
It’s a system where multiple AIs work together. Each agent uses a different tool: one writes code, one runs searches, and another reasons in plain text.
They independently solve the same problem, share their answers, and go through several rounds of refinement until the group reaches a consensus.
The results are stunning. Gemini-2.5 running TUMIX improves performance by up to 17.4% over other reasoning systems, at about half the inference cost. No retraining, no new data, just smarter coordination.

Consider the progress of human civilization. The role of individual geniuses is undeniable, but the power of large-scale group collaboration is even more important.
I’ve always been puzzled by something. Current large models have learned almost all of humanity’s (publicly available) knowledge within a single model. However, this knowledge is often inconsistent and sometimes even contradictory. How do these large models achieve self-consistency?
For example, each of us, despite our limited knowledge, is self-consistent at any given moment. We don’t simultaneously believe in two contradictory ideas.
Individuals form societies and nations, with a division of labor, similar to Darwinian natural evolution. Correct (or viable) ideas triumph over incorrect ones, step by step, driving the progress of the entire society.
The next leap in reasoning might not be a trillion-parameter model, but a network of smaller models learning to think together.