Late last year I was asked to review an AI system that a mid-sized logistics company had built to handle inbound customer queries about shipments. On paper it was working. The single agent behind it could read a tracking number, query the carrier’s API, interpret delay codes, draft a reply and, when the situation warranted, open a support ticket. The team that built it was rightly proud. In a demo it looked close to magic: a customer asked “where is my order and why is it late,” and thirty seconds later a coherent, accurate, faintly apologetic paragraph appeared, complete with a revised delivery estimate.
The problem was that it only looked like magic in the demo. In production, the same system was quietly failing in ways nobody could see. Roughly one query in six produced an answer that was subtly wrong: a delay attributed to weather when it was actually a customs hold, a delivery estimate that ignored a known depot backlog, a refund offered when policy said no refund was due. None of these were catastrophic on their own. Collectively they were eroding trust with the company’s largest accounts, and the operations team had started manually re-checking a growing share of the agent’s output, which defeated the entire point.
When we traced the failures, the pattern was not what the team expected. The agent was not hallucinating in the crude sense. It was being asked to do too many distinct jobs inside a single reasoning pass. Reading a delay code, applying refund policy, judging customer sentiment, and composing a reply are four different competencies, and the more of them we crammed into one prompt and one model call, the more the agent behaved like a talented generalist who has been handed four unrelated tasks and told to finish them all in the next thirty seconds. It did each one adequately and none of them well. What the demo had disguised, and what production exposed, was that the architecture had reached its ceiling. The team’s instinct was to write a longer, more careful prompt. The actual answer was that one agent was no longer enough. This article is about that moment: how to recognise it, and how to move to multi-agent systems without trading one unreliable architecture for a more expensive one.
The Real Problem Is Not Capability, It Is Load
There is a persistent confusion in AI adoption between a model’s capability and a system’s capacity, and it is at the root of most premature multi-agent failures, and of most stubborn single-agent ones.


