The prevailing approach to scaling multi-agent systems has it backwards. Throwing more agents at a problem doesn't reduce complexity, it multiplies it, often by a factor of 17x in error propagation. The hard-won lesson from the field is that a "bag of agents" is not a system; it's a recipe for compounding chaos.
What this means in practical terms is that your current scaling strategy may be working against you. The taxonomy of core agent types, specialists, coordinators, validators, and orchestrators, offers a clearer path. Instead of adding agents indiscriminately, you need to assign distinct roles that limit error cascades. A specialist agent handles one task with precision. A coordinator agent routes work between specialists. A validator agent checks outputs before they pass downstream. An orchestrator agent manages the overall workflow. When each type has a defined function, errors stay contained rather than snowballing across the entire system.
The 17x error trap is not hyperbole. It reflects a structural problem: when every agent can talk to every other agent without guardrails, a single mistake in one node can ripple through the network, amplified at each step. The taxonomy solves this by imposing hierarchy and accountability. Your system stops being a flat, noisy crowd and becomes a structured team. This is the difference between scaling that works and scaling that breaks.
The takeaway is straightforward: stop treating agents as interchangeable units. Assign them roles, enforce boundaries, and measure error propagation at each layer. Your multi-agent system will fail not because you have too few agents, but because you have too little structure. Build the taxonomy first. The agents will follow.