EMPIRICAL COMPARISON OF AGENT COMMUNICATION PROTOCOLS FOR TASK ORCHESTRATION

Abstract

Context. The problem of comparative evaluation of communication protocols for task orchestration by large language model (LLM) agents is considered. The object of study is the process of interaction between LLM agents and external tools, as well as between autonomous LLM agents, during task orchestration.Objective. The goal of this work is to develop a systematic pilot benchmark comparing tool integration, multi-agent delegation, and hybrid architectures for standardized queries at three levels of complexity, and to quantify the advantages and disadvantages in terms of response time, context window consumption, cost, error recovery, and implementation complexity.Method. An open-source health analyzer application was developed to query four public data sources for repository statistics, package downloads, vulnerability data, and community metrics. Thirty queries were run five times for each architecture at three levels of complexity, resulting in 450 fail-safe executions. The same large language model was used in all executions. The statistical analysis used bootstrap confidence intervals, the Mann-Whitney test with Bonferroni correction, and Cliff’s delta effect sizes.Results. A statistically significant crossover effect was found: the tool integration protocol is faster for simple queries, but the multi-agent protocol outperforms on complex multi-project comparisons, consuming fewer tokens through distributed context windows, resulting in lower cost. The hybrid architecture tracks the winner at each level of complexity with near-optimal performance.Conclusions. The scientific novelty is an empirical demonstration that the two protocols exhibit a complexity-dependent performance crossover, showing that they complement each other rather than compete. Context-window bloat in single-agent architectures is quantified for the first time in a pilot case study. The practical significance is a validated decision framework: using a tool integration protocol for single-source queries, a multi-agent protocol for complex orchestration, and a hybrid routing when complexity is unknown at runtime. Directions for further research include extending the benchmark to additional models, distributed deployments, and additional application domains