token-efficiency 2 Benchmark subagent models the way you run them Oct 10, 2026 Taming Qwen Overthinking with GBNF Grammars Apr 28, 2026