The AI Agent That Cost $47,000 While Everyone Thought It Was Working

For eleven days, a multi-agent research system sat in production doing exactly what it was designed to do: agents talking to agents, processing requests, passing messages. Dashboards showed activity. Latency looked normal. No errors fired. The system was healthy. Except it wasn't doing anything useful. Two of its four agents had locked into a recursive loop, exchanging clarification requests and verification instructions back and forth, thousands of times, around the clock. By the time anyone looked at the invoice, the bill was $47,000. Nobody on the team knew until the cloud bill arrived. The Architecture That Looked Right on Paper The system used four LangChain-style agents coordinating via agent-to-agent (A2A) communication to help users research market data. Based on how these architectures are typically structured, the agents likely followed a division of labor similar to this: Research Agent — gathered raw data from external sources Analysis Agent — synthesized findings into stru