miHoYo Burned $300,000 in 13 Hours on a Multi-Agent Experiment That Produced Nothing Usable
The Genshin Impact studio disclosed at Alibaba Cloud Summit that dozens of AI agents reviewing each other's work consumed 47 million yen in tokens and delivered outputs riddled with hallucinations — a cautionary tale for anyone scaling agentic architectures.
miHoYo, the Chinese gaming giant behind Genshin Impact and Honkai: Star Rail, revealed at Alibaba's Cloud Summit that an internal multi-agent experiment burned through roughly $300,000 in API costs in just thirteen hours — and the outputs were effectively unusable. The disclosure, first reported by @FURUYAN123456, has quickly become a reference case for the hidden costs and failure modes of agentic AI systems at scale.
The experiment involved dozens of AI agents arranged in a review-and-feedback loop: agents would generate content, other agents would critique it, revised outputs would be sent back for further review, and so on. The idea — familiar to anyone who has followed the multi-agent hype cycle — was that iterative self-correction would converge on high-quality results. Instead, the system diverged. Token consumption ballooned as agents engaged in increasingly verbose exchanges, and the final outputs were riddled with hallucinations that no layer of automated review managed to catch.
Get our free daily newsletter
Get this article free — plus the lead story every day — delivered to your inbox.
Want every article and the full archive? Upgrade anytime.
No spam. Unsubscribe anytime.