Google’s 260-run study finds multi-agent systems can be 80.8% better — or 70% worse
alex_verem · x · 2026-07-25
Google Research, DeepMind, and MIT ran the largest controlled study on multi-agent systems, testing 260 configurations across OpenAI, Google, and Anthropic models on six benchmarks and five architectures with the same tools and compute budgets.
The key takeaway is task-dependent performance:
- For tasks that split into independent pieces, multi-agent teams beat a single agent by up to 80.8%.
- For sequential tasks where each step depends on the last, every multi-agent setup lost to a single agent, by as much as 70%.
- Uncoordinated agents amplified each other's mistakes 17.2×, while a single coordinator reduced that to 4.4×.
- Once a single agent already reached 45% accuracy, adding more agents tended to produce negative returns.
The paper's practical advice is simple: only go multi-agent when the work can truly be partitioned into independent subproblems.
Related event: Google's Massive Experiments Reveal Multi-Agent Systems' Double-Edged Sword(3 posts)→
More from Research
- Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench — karen_ullrich · 2026-10-27
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Why don't more LLMs ship 10M-token context? A researcher asks about attention tradeoffs — agarwl_ · 2026-10-27
- PSNR below 15 dB in papers likely means the metric was misused, researcher warns — kwangmoo_yi · 2026-10-27
- Ortet's first health eval: 34 open-weight models, 50 tasks, generic rankings fall short — thegautamkamath · 2026-10-27
- Scaling-fit discrepancy: output-head FLOPs are ~65% of compute at the smallest rung, ~3% at 535B — dlwh · 2026-10-27