Asian Journal of Research in Computer Science

Large Language Models (LLMs) are increasingly deployed in multi-agent systems; however, existing frameworks continue to suffer from communication ambiguity, coordination failures, and poor scalability as task complexity increases. This paper introduces COLLAB-LLM, a communication-centric, role-based framework designed to enable reliable and scalable collaboration among LLM agents. The framework combines a structured communication protocol, a hierarchical role architecture, and a dynamic distributed task-graph engine to support coordinated planning, efficient negotiation, and adaptive task execution. COLLAB-LLM is evaluated on over 120 complex, multi-step tasks spanning software engineering, business process automation, and scientific research synthesis. Task success is defined using task-specific completion criteria, with a task considered successful when the aggregate completion score exceeds 0.8. Under identical underlying LLM configurations, COLLAB-LLM achieves an 89% overall success rate, representing a 13–19% improvement over strong single-agent and multi-agent state-of-the-art baselines, with statistically significant gains in performance, communication efficiency, and robustness. Experimental results demonstrate that structured communication and role specialization substantially reduce ambiguity, improve collaboration quality, and enable scalable coordination for teams of up to eight agents. This work establishes foundational design principles for high-performing collaborative AI systems and provides a practical, reproducible pathway toward scalable, human-aligned multi-agent LLM architectures. All experimental artifacts, task definitions, prompts, and evaluation scripts will be released to support reproducibility.