Journal of Advanced Robotics, Autonomous Systems and Human-Machine Interaction
MCP + SKILL: A Deterministic Architecture for Enterprise AI Automation with Selective Intelligence Application
Abstract
Debendra Ray
The proliferation of Large Language Model (LLM)-based autonomous agents in enterprise environments has exposed fundamental architectural limitations including unpredictable execution paths, compounding computational costs, non- deterministic outputs, and complex audit trails that impede regulatory compliance. This paper introduces the MCP + SKILL (Model Context Protocol + Structured Knowledge & Instruction Layer for LLM) architecture, a novel framework that addresses these limitations through deterministic workflow execution with selective intelligence application. The architecture comprises four layers: (1) a lightweight orchestration layer utilizing directed acyclic graphs for state management, (2) SKILL files providing human-readable, version-controlled workflow specifications, (3) Model Context Protocol servers offering standardized tool interfaces, and (4) external system integrations. The key contribution lies in the formalization of “decision points”—explicitly defined workflow stages where LLM reasoning provides genuine value— and a conditional routing mechanism that constrains LLM invocation to these points. Empirical evaluation through a B2B customer onboarding case study demonstrates a 97.9% reduction in LLM API costs, 85.3% reduction in execution latency, and 100% output determinism for identical inputs compared to autonomous agent baselines. The framework achieves enterprise-grade auditability while preserving intelligent capability for genuinely ambiguous scenarios. We position MCP + SKILL as complementary to emerging AI governance frameworks, addressing the execution layer while governance frameworks address authorization.