There's an anecdotal paper 'How We Broke Top AI Agent Benchmarks: And What Comes Next' https://moogician.github.io/blog/2026/trustworthy-benchmarks...