TCP-MCP: Landscape-Guided Co-Evolution of Prompts and Communication Topologies for Multi-Agent Systems

TCP-MCP introduces a co-evolution framework that simultaneously optimizes AI agent prompts and communication network topologies, achieving state-of-the-art accuracy on multiple benchmarks while reducing token consumption by up to 5.69x compared to existing multi-agent systems. The approach treats prompt design and communication structure as interdependent variables rather than independent parameters, offering a practical methodology for cost-efficient multi-agent AI system design.

TCP-MCP addresses a fundamental design problem in multi-agent systems: the interdependence between how agents are instructed (prompts) and how they exchange information (topology). Traditional approaches optimize these elements separately, overlooking how an agent's interpretation of information depends on its directives. This framework uses a landscape probe at initialization to guide early search behavior and Pareto-front diagnostics to balance three competing objectives: accuracy, computational cost, and structural simplicity. The research demonstrates significant practical value by achieving competitive or superior performance on standardized benchmarks—MMLU-Pro (82.66%), MMLU (89.96%), and GSM8K (96.61%)—while dramatically reducing token usage, a critical metric for production deployment costs. This work reflects a broader maturation in multi-agent AI systems, moving beyond monolithic or randomly-structured approaches toward principled, co-optimized designs. The token efficiency gains are particularly important as large language model costs scale with inference volume; a 5.69x reduction represents substantial operational savings for organizations deploying multi-agent reasoning systems. The methodology's reliance on Pareto optimization enables practitioners to select operating points matching their specific constraints—prioritizing accuracy for critical applications or cost-efficiency for resource-constrained scenarios. As multi-agent systems transition from research novelties to production infrastructure, frameworks like TCP-MCP that jointly optimize multiple system dimensions will likely become industry standards, influencing how organizations architect reasoning systems.

• →TCP-MCP co-evolves agent prompts and communication topologies as unified variables, outperforming independently-optimized approaches across three benchmarks.
• →The framework achieves up to 5.69x token efficiency improvements compared to debate-style multi-agent systems at equivalent accuracy levels.
• →Pareto-front optimization enables three-way balancing of task performance, computational cost, and structural complexity for different deployment scenarios.
• →Results suggest joint optimization of prompts and communication structure is more effective than optimizing either component in isolation.
• →The approach uses landscape probing at initialization to calibrate search behavior and guide exploration efficiency.