Document Type

Article

Abstract

There is growing interest in automating business activities with Agentic Artificial Intelligence (AI) due to latter's seeming ease of use. However, little is known on when they are suitable for a task over other alternatives developed over the years - local computation, REstful State Transfer (REST), and Simple Object Access Protocol (SOAP) - considering development speed, performance, and operational cost. We explore this with a small mathematical task evaluating five methods for automated mathematical expression evaluation across a benchmark of 1,000 equations where semantics of operator precedence has to be preserved. We ran this setup across a native Function Calling baseline (python library), REST and SOAP microservice architectures, and Agentic systems with two large language model (LLM) backends (Groq cloud API and a three locally run Ollama instances). We find that deterministic methods achieve 100% accuracy with negligible latency, while Agentic approaches that use Large Language Models reach at most 78.6% accuracy at roughly 30,500x the computational cost and 72,000x the latency. Our results demonstrate that LLMs are fundamentally unsuitable for exact arithmetic evaluation and motivate an informed architectural decision framework (ADF) based on accuracy and throughput requirements. We use our results to propose a decision framework that can help decide which architecture to use based on tradeoff considerations. We recognize that an actual interaction with an agent may span both reasoning and linguistic tasks. Hence our recommendation is to explore multiple architectural choices (based on tasks type) rather than one-size fits all (LLMs).

Code is available at - https://github.com/Ritvik-G/adf/

Rights

© 2026, The Authors

Download

Share

COinS