A Five-Engine Reference Architecture for Cost-Aware, Self-Adaptive Agentic AI: An Engineering Case Study in the Internet of Value

This paper is about the engineering, not the algorithms, of an agentic risk-intelligence platform running on the blockchain Internet of Value. Its question is how to build one AI agent that stays cheap to run and stays calibrated for months under adversarial inputs.

It makes three contributions: a five-engine reference architecture, a cost-aware engine that tracks the inference, trust and action cost of every decision, and a self-adaptive loop that recalibrates on a versioned manifest.

Two operational case studies provide the evidence, and the paper closes with four falsifiable engineering propositions.