Large Language Models (LLMs) are increasingly used for code generation, including complex distributed systems. However, their reliability in stateful and failure-prone environments remains unclear. In this paper, we present a structured evaluation framework for assessing LLM-generated implementations of distributed protocols, including Two-Phase Commit, Ring Election, and Raft. We introduce a simulation environment that injects realistic network faults such as message loss, delay, and duplication. Our results show that while LLMs perform well on simpler protocols and code completion tasks, they struggle with complex consensus algorithms and exhibit inconsistent debugging behavior. We further analyze the relationship between model size and performance, finding that parameter count alone does not predict effectiveness. These findings highlight the need for systematic testing and verification in agentic code generation systems.