Developing control software for programmable logic controllers is labor intensive, highly domain specific, and tightly coupled with proprietary engineering tools. While Large Language Model-based code generation has already improved productivity in conventional software engineering, transferring these gains to automation engineering is more demanding because PLC projects combine code generation with hardware configuration, vendor-specific tool interaction, and strict correctness requirements. In addition, known LLM limitations such as prompt bloat, non-determinism, and insufficient tool knowledge directly affect engineering quality.
This paper presents a systematic benchmark of seven agentic architectures for PLC engineering, ranging from plain LLM prompting to MCP-enabled, single-agent, and multi-agent workflows with Retrieval-Augmented Generation and Agent2Agent communication. Using two industrial laboratory automation use cases, the benchmark demonstrates end-to-end generation of compilable and executable automation projects and quantifies functional correctness, tool usage, token consumption, and cost. The results show that standardized tool access and explicit closed-loop orchestration substantially improve reliability and that, in the investigated setting, the single-agent architecture provides the strongest correctness-resource trade-off, whereas more complex multi-agent setups improve modularity and openness but introduce additional coordination overhead without consistent correctness gains. Building on these findings, the paper further discusses the infrastructure, information-layer, and OT-security conditions required to move from isolated code generation toward governable agentic automation workflows which might span multiple automation engineering lifecycle phases.