Google Research has published a blog post calling for a more rigorous, scientific understanding of scaling agent systems. The post centers on a deceptively simple question: when and why do agent systems actually work?
The authors argue that agentic designs should not be treated as a default win. Instead, they advocate for systematic study of the factors that make multi-agent setups beneficial, such as task structure, inter-agent communication, and the cost of additional steps.
This framing positions agent systems as an engineering pattern that needs its own scaling laws and failure analysis, much like large language models have received. The post is an early step toward a more predictive theory of agentic AI, rather than relying on anecdotal gains.