An Experimental Comparison of Multi-Agent and Monolithic Prompt Architectures Across LLM Scales for Hardware Diagnostics

• João P. A. Lima UFC
• Carlos Caminha UFC
• Victor A. E. Farias UFC
• João P. P. Gomes UFC
• Francisco L. F. Pereira UFC
• Maria de Lourdes M. Silva UFC
• Javam Machado UFC
• José S. Andrade Jr. UFC

DOI: https://doi.org/10.5753/eniac.2026.31096

Resumo

Monolithic prompting is a common baseline but struggles in complex multi-task settings, especially with smaller models, whereas multi-agent approaches decompose workflows into specialized prompts. In this study, we compare both across 26 open-source LLMs in a hardware diagnostics chatbot, using a benchmark covering guardrails, diagnostic reasoning, and constrained dialogue, with a genetic algorithm-based prompt optimizer to ensure a fair comparison. Results show that multi-agent benefits depend on task complexity and model scale, improving consistency and average performance primarily for smaller models on complex tasks, suggesting that task decomposition can mitigate the limitations of smaller LLMs under controlled conditions.

Referências

Caminha, C., Silva, M. d. L. M., Chaves, I. C., Brito, F. T., Farias, V. A., and Machado, J. C. (2025). Evaluating llms and prompting strategies for automated hardware diagnosis from textual user-reports. In 2025: Anais do LII Seminário Integrado de Software e Hardware.

Caminha, C., Silva, M. d. L. M., Chaves, I. C., Brito, F. T., Farias, V. A. E., and Machado, J. C. (2026). Diaghw: A compact llm for hardware failure diagnosis via a novel knowledge distillation pipeline. In de Freitas, R. and Furtado, D., editors, Intelligent Systems, pages 393–407, Cham. Springer Nature Switzerland.

Cui, W., Zhang, J., Li, Z., Sun, H., Lopez, D., Das, K., Malin, B. A., and Kumar, S. (2024). Phaseevo: Towards unified long-context prompt optimization for large language models. In First Workshop on Long-Context Foundation Models@ ICML 2024.

Dettmers, T. and Zettlemoyer, L. (2023). The case for 4-bit precision: k-bit inference scaling laws. In Proceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org.

Dushkin, R. V., Klimov, V. V., Anzin, Y. A., Dushkin, K. R., and Zvorygin, V. A. (2026). Cost-efficient distributed ai architectures: Lightweight llm agents vs gpt-4o monoliths in academic workflows. In Samsonovich, A. V., Ramos, F., and Liu, T., editors, Biologically Inspired Cognitive Architectures 2025, pages 181–190, Cham. Springer Nature Switzerland.

Guo, Q., Wang, R., Guo, J., Li, B., Song, K., Tan, X., Liu, G., Bian, J., and Yang, Y. (2023). Connecting large language models with evolutionary algorithms yields powerful prompt optimizers. arXiv preprint arXiv:2309.08532.

Kurt, U. (2026). Which quantization should i use? a unified evaluation of llama.cpp quantization on llama-3.1-8b-instruct.

Liu, P. R., Bansal, S., Dinh, J., Pawar, A., Satishkumar, R., Desai, S., Gupta, N., Wang, X., and Hu, S. (2025). Medchat: A multi-agent framework for multimodal diagnosis with large language models. In 2025 IEEE 8th International Conference on Multimedia Information Processing and Retrieval (MIPR), pages 456–462. IEEE.

Schulhoff, S., Ilie, M., Balepur, N., Kahadze, K., Liu, A., Si, C., Li, Y., Gupta, A., Han, H., Schulhoff, S., et al. (2024). The prompt report: a systematic survey of prompt engineering techniques. arXiv preprint arXiv:2406.06608.

Wu, C., Lin, Z., Fang, W., and Huang, Y. (2023). A medical diagnostic assistant based on llm. In China Health Information Processing Conference, pages 135–147. Springer.