AI swarms, like the 2,000 agents used in Cursor’s FastRender project, can produce functioning code but struggle with coherence. Each agent operates independently, often solving the same problems with different solutions, leading to a lack of consistency. For example, two agents might choose different HTTP clients, which complicates the overall system. Even if both choices work, the variability means the swarm lacks reproducibility and reliable outcomes. The issue deepens when multiple models are involved. Different AI models, such as GPT-5.2 or Claude Opus 4.2, yield divergent answers even for the same prompt due to variations in training data and preferences. This divergence means agents are not merely uncoordinated; they consult different “oracles” with their own biases. Even when using the same model, factors like floating-point arithmetic and hardware differences can lead to inconsistent outputs, compounding the problem. Architecture requires global coherence and adherence to invariants, such as consistent error handling and dependencies. Swarms, in contrast, optimize locally without awareness of the entire system. They lack the mechanisms to enforce architectural constraints, as each agent works in isolation and does not retain context from prior decisions. As the number of agents increases, the likelihood of accidental coherence diminishes significantly, making it nearly impossible for swarms to produce a coherent architectural framework. The article emphasizes that this isn't a mere technical glitch but a fundamental mismatch between the nature of swarms and the requirements of architectural design.