Towards a science of scaling agent systems: When and why agent systems work

Google Research just put a dent in one of AI's favorite talking points: that more agents always means better results. In a new paper called "Towards a Science of Scaling Agent Systems," the team ran 180 different agent configurations across five architectures — a lone single-agent setup plus independent, centralized, decentralized, and hybrid multi-agent teams — and tested them on four benchmarks covering finance, web browsing, planning, and tool use, using GPT, Gemini, and Claude models.

The headline finding is that architecture choice matters as much as raw model power, and the two interact in ways that punish sloppy design. On tasks that break cleanly into parallel chunks, like financial analysis where one agent can chew on revenue while another checks costs, a centralized orchestrator boosted accuracy by 80.9% over a single agent. But flip to something sequential, like the planning benchmark PlanCraft, and every single multi-agent variant lost ground — accuracy dropped between 39% and 70%. The agents spent their effort coordinating instead of thinking, leaving what the researchers call a shrunken cognitive budget for the actual problem.

There's also a tool-use penalty. Once a task demands access to 16 or more tools, the cost of keeping multiple agents synced starts outpacing whatever benefit they add. And the safety angle is arguably the most useful part of the paper: independent agents working without any communication amplified errors 17.2 times over, since nobody was checking anyone's work. Centralized systems, where an orchestrator reviews outputs before finalizing them, kept that amplification down to 4.4x — acting as a kind of built-in error trap.

To turn all this into something practitioners can actually use, the team built a predictive model that looks at task properties like tool count and how decomposable a job is, then recommends an architecture. It's not perfect — an R² of 0.513 — but it picked the right setup 87% of the time on tasks it hadn't seen before. The bigger implication is that as models like Gemini keep getting stronger, that doesn't make multi-agent systems obsolete; it makes matching architecture to task more urgent, not less.