Abstract

Large language model (LLM) agents have moved from single-turn assistants to persistent, tool-using systems that plan, delegate, and cooperate. This shift has produced an explosion of multi-agent system (MAS) frameworks, interoperability protocols, and orchestration patterns, yet production deployments continue to report failure rates far higher than single-agent baselines. This paper synthesizes the current state of AI agents and multi-agent systems as of mid-2026, organizes the coordination design space into a four-part topology taxonomy, and reviews the emerging empirical literature on why multi-agent systems fail. Building on documented gaps, particularly the absence of adaptive, trust-aware coordination in widely used frameworks. We propose Trust-Weighted Adaptive Consensus (TWAC), a coordination-layer mechanism that combines capability-aware sparse shortlisting, an exponentially-weighted per-agent trust ledger, and uncertainty-gated verification. We evaluate TWAC against round-robin and centralized-auction baselines in a controlled simulation across 500 tasks and 30 independent trials. TWAC improves task success rate while reducing communication overhead relative to full-broadcast coordination and lowering cascading-failure rates by more than an order of magnitude. We conclude with future research implications for trust-based coordination in agentic AI systems.

References

• Lang Chain, State of AI Agents Report, 2025.
• McKinsey & Company, The State of AI in 2025: Agentic AI Adoption Survey, 2025.
• Gartner, Forecasts on Enterprise Agentic AI Adoption and Project Cancellation Rates, Gartner Research, 2025, 2026.
• Finin, T. W., Fritzson, R., McKay, D. P., & McEntire, R, KQML as an Agent Communication Language. Proceedings of the Third International Conference on Information and Knowledge Management (CIKM ’94), Pages: 456–463, 1994.
• FIPA, Communicative Act Library Specification, Foundation for Intelligent Physical Agents, 2002.
• Cao, X, “Coordination as an Architectural Layer for LLM-based Multi-Agent Systems”. arXiv:2605.03310, 2026.
• Wu, Q, “AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation”, Microsoft Research, 2023.
• Ehtesham, A., Singh, A., Gupta, G. K., & Kumar, S, “A survey of agent interoperability protocols: Model Context Protocol (MCP), Agent Communication Protocol (ACP), Agent-to-Agent Protocol (A2A), and Agent Network Protocol (ANP)”, arXiv:2505.02279, 2025.
• Cemri, M., Pan, M. Z., Yang, S., Agrawal, L. A., Chopra, B., Tiwari, R., Keutzer, K., Parameswaran, A., Klein, D., Ramchandran, K., Zaharia, M., Gonzalez, J. E., & Stoica, I, “Why do multi-agent LLM systems fail”, arXiv:2503.13657, 2026.
• Wynn, T, Debate Dynamics and Dissent Suppression in Multi-Agent LLM Deliberation, 2025.
• Xia, C, “Are more LLM Agents Always Better? A Comparative Study of Single and Multi-Agent Performance”, arXiv preprint, 2024.
• Xu, Y, “Revisiting the Necessity of Multi-Agent Decomposition for LLM Task Solving”, arXiv preprint, 2026.
• Pappu, A., El, B., Cao, H., di Nolfo, C., Sun, Y., Cao, M., & Zou, J, “Multi-Agent Teams Hold Experts Back”, arXiv:2602.01011, 2026.
• Kim, J, “Scaling Laws for Multi-Agent LLM System Architectures”, 2025.
• S. Banumathi, "IOT SENSORS AND APPLICATIONS: A SURVEY", International Journal of Scientific & Technology Research, 2019.
• P Subhasri, "Bipolar Disorder Detection Using Machine Learning Technique", Proceedings of 4th International Conference on Mathematical Modeling and Computational Science: ICMMCS 2025, Volume 3, Pages: 458 - 470, 2025.
• V. Jayaraj, "Topology Discovery and Modified Tabu Search for Efficient Routing in Wireless Ad Hoc Networks", Indian Journal of Science and Technology, Vol 7(12), 2088–2095, December 2014.