FHIR-AgentBench: Benchmarking LLM Agents for Realistic Interoperable EHR Question Answering