Doing Less with More: A First-Principles Exploration of the Suitability of Agentic Computing over Alternative Architectural Choices

Introduction

The recent surge of interest in Agentic Artificial Intelligence—systems where large language models (LLMs) orchestrate tools and services—promises easier automation of business workflows. Yet the hype masks a critical question: When does an agentic approach actually outperform traditional architectures such as native function calls, REST, SOAP, or the Model Context Protocol (MCP)? Garimella, Srivastava, and Sheth address this by benchmarking a simple yet representative problem: evaluating 1,000 arithmetic expressions while preserving operator precedence. Their study compares deterministic methods, direct LLM inference, and agentic pipelines that invoke MCP‑based tools.

Deterministic Baselines Shine in Speed and Accuracy

Three deterministic architectures—native Python function calls, a REST microservice, and a SOAP endpoint—achieved 100 % correctness with negligible latency (on the order of microseconds). These results confirm that for tasks requiring exact computation, classic procedural or service‑oriented designs remain the gold standard. The MCP microservice, while slightly more complex, also delivered perfect accuracy with only modest overhead compared to the bare function call.

Direct LLM Evaluation Falls Short on Precision

When the same arithmetic workload was sent directly to LLMs via a Groq cloud API and five locally hosted Ollama models, accuracy dropped to a maximum of 92.9 %. More strikingly, latency ballooned to 22,000 × that of the native baseline. This gap underscores a fundamental limitation: LLMs excel at probabilistic language tasks but are ill‑suited for deterministic numeric reasoning without external tooling.

Agentic Systems Bridge the Gap—At a Cost

The authors introduced agentic pipelines where local LLMs invoke MCP‑exposed tools to perform the arithmetic. The best‑performing agent reached 99.9 % accuracy, nearly matching deterministic methods. However, the latency incurred was roughly 1,000,000 × the baseline, reflecting the overhead of multiple model‑tool interactions, network hops, and orchestration logic.

Practical Guidance: An Architectural Decision Framework (ADF)

From these empirical findings, the paper proposes an Architectural Decision Framework that balances two primary dimensions:

1. Accuracy Requirements – Tasks demanding exact results (e.g., financial calculations, scientific data processing) should favor deterministic services.
2. Throughput & Latency Constraints – Real‑time or high‑volume scenarios benefit from low‑overhead APIs like REST or native calls.

When a workflow involves both precise computation and higher‑level reasoning or natural‑language interaction, a hybrid approach may be optimal: use deterministic services for the numeric core and reserve agentic components for surrounding conversational or decision‑making steps.

Closing Thoughts

The study makes clear that agentic AI is not a universal replacement for established architectural patterns. Its strength lies in flexibility and the ability to integrate heterogeneous tools, but this comes with substantial latency penalties for exact tasks. Practitioners should evaluate the accuracy‑throughput trade‑off early in the design phase, selecting the simplest architecture that meets the required guarantees. By doing so, organizations can avoid the costly pitfall of “doing less with more” and instead harness the right mix of deterministic services and agentic intelligence where each adds genuine value.