Large Language Model (LLM)-based AI agents have advanced beyond simple text generation to perform real-world tasks using external tools. However, there remains no systematic reference model for analyzing agent workflows. This paper proposes Pylon-7, a 7-layer reference model inspired by the OSI model that decomposes AI agent workflows into seven independent layers, positioning the Model Context Protocol (MCP) as the inter-layer gateway. We conducted 615 experimental runs (465 main + 50 L2.5 ablation + 100 L3 component decomposition ablation) across 10 infrastructure operation scenarios at 5 MCP depth levels (L0-L4) using Qwen 2.5 7B and GPT-OSS 20B models on commodity hardware (CPU-only, no GPU) to explore the impact of MCP layering on efficiency (token cost), accuracy (task quality), and safety (privilege control). Key findings: (1) MCP layering (L0→L3) reduced tokens by 47% while improving accuracy by 37%. (2) L3 (structured output + candidate actions) is the sweet spot. (3) A small model (7B)+MCP combination was 3.5x cheaper and 14.4%p more accurate than a large model (20B) alone. (4) Tasks practically impossible without MCP achieved perfect performance with MCP. (5) Above L3, both 7B and 20B models achieved 0.93+ accuracy, suggesting MCP structure may dominate over model size.
Hyunwoo Jeon (Sat,) studied this question.