Enterprises once relied on bare functional calls for routine automations, but companies increasingly deploy coordinated language model agents to execute long, branching standard operating procedures across corporate workflows. Business leaders often adopt these multi-agent designs under the assumption of universal superiority, yet the underlying computational spend rarely delivers a proportionate commercial return.
Escalating architectural complexity frequently fails to justify steep increases in computing costs and operational delays across enterprise processes. A research team led by Subhrangshu Nandi, Hitesh Arora, Yunfei Bai, Udita Patel, Hannah Kwon, Wei Zhang, Weian Chen, Chloe Ruoqi Li, Alberto Sconfienza, Ayush Goyal, Yasmine Lee, Ming Xiong, and Prathap Ramachandra at Amazon, alongside Chirag Shah affiliated with Amazon and the University of Washington, ran a controlled study evaluating agent performance across 13 enterprise SOPs. Holding the model claude-opus-4-8, tasks, and tools fixed, the researchers evaluated four designs structured as eight experimental arms: a function-calling single agent, a skill-augmented agent, centralized multi-agent orchestration, and a triage swarm. To ensure environment stability, SOP-Bench features deterministic mock-tool dispatch that fixes tool outcomes so that observed differences reflect architecture and skill design.
The Realized Cost of Multi-Agent Coordination
The financial footprint of orchestration mechanisms alters the unit economics of business workflows. The researchers found that every coordinated arm spends 1.4 to 4.7 times the tokens of the single agent, alongside 2.6 to 5.8 times the latency.
"Every coordinated arm (multi-agent or skill-augmented) spends 1.4–4.7× the tokens and 2.6–5.8× the latency of the single agent for gains that are usually small and rarely significant, making the single agent the efficient default"
This resource consumption means that orchestrator calls, agent handoffs, and repeated context transfers generate immediate operational expenses that standard productivity gains struggle to offset. Individual procedural implementations expose this economic divergence. Deploying an elaborate swarm to that operational track burns cash and compute while degrading execution quality. Architectural payoff is strictly SOP-dependent rather than a general platform advantage.
Implementation Quality vs Architectural Complexity
A counterargument suggests that added architectural complexity remains unavoidable for intricate enterprise procedures despite high token and latency overheads. The data indicates that structural complexity alone does not solve the underlying procedural execution problem. The authors introduced a skill-definition quality rubric to build matched high- and low-quality skill corpora, demonstrating that execution success hinges on implementation artifacts rather than orchestrator schemes.
"A plausible multi-agent implementation loses to the single agent, yet a targeted redesign of it becomes the only arm to significantly beat the baseline ( 75.6% vs.67.4%)."
High-quality skills lift task success to 73.1% from 67.4%, whereas degraded skills fall below that baseline. Simply switching to a multi-agent framework produces losses against the standard single agent, whereas targeted structural adjustments enable higher completion rates. If a plausible multi-agent design trails a baseline single agent until engineers undergo targeted workflow refactoring, can enterprise software teams justify absorbing a 4.7-fold increase in tokens and a 5.8-fold latency penalty on claude-opus-4-8 across their standard operating procedures?
What doesn’t add up
A fourfold jump in token consumption for a marginal performance gain marks the exact boundary where enterprise automation hits a wall. The architectural urge to stack multiple agents simply collapses without rigorous component engineering, as extra coordination layers amplify response latency without guaranteeing task completion. Orchestration complexity ends up acting as an expensive cover for mediocre skill design, leaving companies to subsidize computational overhead instead of capturing actual value.