Per-Agent Cost Tracking Exposes Silent Overspending in Multi-Agent AI Systems

Multi-agent AI systems on AWS can silently overbill users by roughly 40% even when performance dashboards show no errors or latency issues. A 2025 research paper (MAST, arXiv:2503.13657) analyzed 150 traces across seven multi-agent systems and found failure rates between 41% and 86.7%, with many failures completing successfully on the surface. Traditional application monitoring tools are insufficient for AI agents because they do not track token usage, reasoning cycles, or per-step costs. A developer built a read-only AWS Account Investigator agent crew to demonstrate how per-agent cost data can be wired into trace spans to detect these hidden waste patterns. The project is designed to run locally at no cost without modifying any AWS account resources.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in