More agents help parallel work. They hurt sequential work.
Google Research tested 180 architecture configurations and found multi-agent systems aren't a general upgrade — whether they help or hurt depends on the shape of the task, not the number of agents thrown at it.
The task shape decides the answer, not the agent count
Google Research ran the same kind of question we get asked constantly — will more agents make this better — across 180 configurations and five different architectures, and found the honest answer is: it depends what the work actually looks like. Tasks that break cleanly into independent pieces run in parallel saw large gains from multiple agents working at once. Tasks that require one continuous chain of reasoning — planning work, essentially — got substantially worse under the same multi-agent setups, because splitting a single train of thought across agents fragments it instead of speeding it up.
More agents also means more ways to be wrong
The less-discussed cost of adding agents is that mistakes compound faster than capacity does. Agents working independently amplified errors far more than agents whose output passed through some form of central check before being accepted. The practical implication isn't 'avoid multiple agents' — it's that a multi-agent system without a validating layer isn't just adding throughput, it's adding failure surface. The orchestrator that checks outputs before they combine is what keeps the error rate from compounding as fast as the output does.
Diagnose the shape of the work before you pick the architecture
The researchers found they could predict, with high accuracy, which architecture would perform well just from measurable properties of the task itself — how decomposable it is, how many tools it touches. That's the useful discipline for anyone deciding whether a workflow needs one agent or several: work out whether it's actually parallel or actually sequential before committing to an architecture, rather than treating 'more agents' as a default upgrade path.
The architecture is a consequence of the task, not a starting assumption. Diagnose the shape of the work before deciding how many agents it needs.
Source perspective: Google Research ↗. This AventeqAI brief is an original synthesis and commentary, not a reproduction of the source article.