Response time is a million times slower: scientific study reveals the cost of automating calculations with agents

When organizations rush to replace traditional software systems with AI agents, most assume that the flexibility of generative models compensates any difference in computational performance. However, a research paper we examined, accepted at the AI Advancement Association’s Autumn Conference on Trustworthy Agent Systems, put this assumption to a rigorous test, revealing that attempting to perform simple mathematical tasks via agents may cost the system a delay of roughly one million times compared with direct programmatic calls.

The study was prepared by researchers Ritvik Garimela, Biblav Srivastava and Amit Sheth under the title “Achieving Less by Relying on More: A First-Principles Exploration of the Suitability of Agent-Based Computing Compared with Alternative Architectural Options.”The team compared seven different architectural approaches using a benchmark consisting of one thousand arithmetic expressions to test how well the precedence semantics of mathematical operations are preserved.The comparison included four conventional deterministic solutions: the Python library for native programmatic calls, a micro-service architecture based on REST interfaces, a simple object access protocol, and a micro-model-context protocol architecture. In addition to these solutions, the researchers evaluated direct inference of large language models via a Grok cloud interface and five local models on the Olama platform, as well as agent systems that invoke model-context protocol tools across five local models.

The results showed that the deterministic methods achieved 100 % accuracy with a negligible response time. In contrast, direct evaluation of language models recorded a maximum accuracy that did not exceed 92.9 %, but it incurred a latency exceeding 22 000 times that of the baseline programmatic call.When the models were integrated into agent systems that invoke specialized tools, accuracy rose to 99.9 %, but response time jumped to roughly one million times that of the reference software architecture.This gap demonstrates that language models are unsuitable for strict mathematical processing, prompting the researchers to propose an architectural decision-making framework that identifies the most appropriate technical option based on accuracy requirements and traffic volume, rather than resorting to a single architecture that relies entirely on language models.

These figures have a direct impact on technical decision-making rooms in the Gulf, Egypt and the Levant, where banks, fintech firms and government agencies race to build smart agents for daily operations. If a technical team in Riyadh, Cairo or Dubai decides to delegate tasks such as computational matching, invoice settlement or fee calculation to a linguistic agent that invokes external tools on each transaction, the organization will face a massive jump in computing costs and cloud consumption, as well as severe latency that disrupts concurrent transactions requiring instantaneous response times.

Infrastructure efficiency requires a decisive separation between context comprehension and deterministic operation execution.System engineers and digital-transformation leaders should keep precise calculations and data matching within established APIs and mature codebases, limiting the role of smart agents to linguistic inference and unstructured interaction management. This balance protects the operating budgets of regional organizations and ensures that companies do not pay a high price for complex technologies that perform tasks that simple code lines can accomplish with full accuracy and negligible delay.