Researchers at Amazon published a study on AI agent architectures, titled Agent Architecture Cost–Success Trade-offs: A Controlled Evaluation across Enterprise-Style SOPs, in November 2026. The team evaluated 13 enterprise workflows to compare the performance of single-agent systems against multi-agent setups.
The baseline single-agent model achieved a task success rate of 67.4%. Adding high-quality skills improved the success rate to 73.1%, while low-quality skills reduced it to 59.8%. Multi-agent configurations showed varied results, with the lowest scoring 58.2% and the highest reaching 75.6%.
Complex multi-agent systems consumed between 1.4 and 4.7 times more tokens than the single-agent baseline. Additionally, latency increased by 2.6 to 5.8 times. The researchers noted that no architecture provided a statistically significant improvement across all 13 tested workflows.
The study concludes that implementation quality is a more critical factor than the chosen architecture. Amazon recommends that companies start with a single-agent model and only increase complexity if a specific workflow demonstrates a clear, measurable benefit.
