A new survey paper published by IEEE reveals a critical insight into the architecture of modern multi-agent systems: two-thirds of their effectiveness stems not from the underlying large language models, but from the supporting infrastructure. This framework, developed by control systems researchers, systematically maps the evolution from classical multi-agent coordination to current LLM-driven approaches, analyzing everything from swarm intelligence to device-edge-cloud deployment. The study challenges the prevailing assumption that the reasoning capabilities of the AI model are the primary driver of success. Instead, it highlights that the majority of value lies in the "harness"—the three-layer optimization architecture surrounding the model. These layers include the knowledge layer, which manages memory and context, and the system layer, which handles deployment topology and routing logic. The paper notes that while the model layer is essential for generating intelligence, it represents only one-third of the optimization surface. In practice, this means that when a multi-agent system fails, the root cause often resides not in the model's lack of reasoning, but in the surrounding engineering. Common issues such as instruction fade-out, context overflow, and error recovery loops are frequently resolved by adjusting the harness rather than retraining the model. This pattern holds true across single-agent and multi-agent systems, with the ratio of harness dependency becoming even more pronounced as systems scale and interact. For developers building these systems, the paper advocates for a shift in strategy. First, teams should avoid over-investing in model-level tuning, such as prompt engineering or fine-tuning, which constitutes only a small fraction of the work. The focus must expand to the knowledge and system layers, including robust memory architectures, efficient context management, and sophisticated routing logic. Communication between agents is identified as a significant bottleneck. The survey documents emerging protocols like MCP, A2A, and ANP, while emphasizing the need for pruning strategies to reduce communication costs. Engineers should treat inter-agent communication as a first-class problem, designing it with the same rigor as the model itself. Furthermore, the architecture must account for heterogeneity. Since different models operate at varying scales and latencies, the harness must route tasks intelligently. A key example is the confidence-aware escalation pattern, where low-confidence inference on an edge device triggers assistance from a more powerful cloud model. This is a design solution, not a training technique. Finally, systems should be planned for co-evolution, ensuring that interfaces are ready to coordinate LLM-based planners with classical controllers, which often present the most difficult integration challenges. The paper concludes with a clear distinction between science and engineering. While the model layer is the domain of scientific research, the knowledge and system layers are the realm of engineering. In production environments, engineering wins. The model provides the raw intelligence, but the harness constructs the functional system. Ultimately, the consensus is that two-thirds of multi-agent intelligence is harness.