We propose a benchmark to evaluate the large language models' instruction following ability in agentic scenarios.