Multi-AI agent tokens up to 4.7x more… Amazon researchers propose single configuration

Amazon and University of Washington researchers, in a paper released November 7, recommend a single-agent setup as the default starting point for enterprise AI agents after testing four agent architectures and eight configurations across 13 standard operating procedures using Claude Opus 4.8 with identical tasks and tools. The highest-performing multi-agent configuration achieved the best task success rate but consumed 4.7x more tokens and 3.1x more processing time than the single-agent baseline. The finding matters because enterprises scaling agentic workflows face compounding inference costs and latency that can outweigh accuracy gains from multi-agent designs. The study held the model, tasks, and tools constant, isolating architecture as the only variable, which gives the token and time multipliers a clean basis for cost comparisons. What to watch next is whether vendors and enterprises adopt single-agent defaults for routine SOP automation and reserve multi-agent orchestration for tasks where the success-rate premium justifies the added token and latency burden.