Google study across 260 agent setups shows multi-agent gains depend on task structure

alex_verem · x · 2026-07-25

Google Research, DeepMind, and MIT published a large controlled study of multi-agent systems, titled Towards a Science of Scaling Agent Systems. The team evaluated 260 configurations across three model families, six benchmarks, and five architectures while holding tools and compute constant to isolate the effect of coordination.

Main findings:

Parallelizable tasks benefit from more agents. On workloads that can be split into independent pieces, multi-agent systems improved performance by up to 80.8%.
Sequential tasks get worse with more agents. When each step depends on the previous one, multi-agent setups underperformed a single agent by as much as 70%.
Coordination matters a lot. Uncoordinated teams amplified errors 17.2×; adding a coordinator reduced the multiplier to 4.4×.
There is a ceiling. Once one agent already reaches about 45% accuracy, adding more agents tends to produce negative returns.

The paper argues that the common mistake is scaling multi-agent systems without checking whether the task structure actually supports independent decomposition.

Original post →

Google study across 260 agent setups shows multi-agent gains depend on task structure

More from Research