Towards AI-Agent-Accelerated Research in High-Energy Physics: A Case Study in Monte Carlo Analysis

We investigate how far contemporary LLM-based agents can participate in multistep computational high-energy-physics research under reproducible human supervision. Using a previously published CEPC Bell-observable analysis in $e^+e^- \to ZH$, $Z \to q\bar q$, $H \to ZZ^* \to 4μ$ as a case study, we test two capabilities. First, the agent performs an instruction-constrained reproduction: it is explicitly instructed not to use the published numerical results as analysis targets, although no technical information-access barrier is imposed. The analysis definition is fixed before the formal comparison, and the resulting detector-level $I_3$ central values reproduce the four published benchmark points to within about 2% without retuning to those values. Second, the agent is used for an open-ended extension with no target answer supplied. Using 200,000 fast-simulated events, a stage-resolved audit identifies muon isolation as the largest single contribution to the hard-to-detector shift, and subsequent diagnostics connect the effect to angle-dependent isolation efficiency, proximity to the hard $Z \to q\bar q$ system, and mixed electromagnetic activity from muon final-state radiation and hadron decays. Throughout, the agent assists with workflow planning, code and configuration inspection, Monte Carlo production, anomaly diagnosis, control construction, statistical analysis, and result organization, while state-changing actions and final interpretation remain under human control. The study shows that the value of AI agents in HEP may lie not only in faster code generation, but in faster scientific iteration, and motivates systematic benchmarking across historical HEP analyses using correctness, human intervention, and researcher-time efficiency as explicit metrics.