From Prompt Engineering to Agent Engineering: A Layered Stack for Protocols, Workflows, and Declarative Intent in Agentic AI

Large language models are moving rapidly from conversational interfaces toward agentic systems that plan, invoke tools, maintain task state, and coordinate across software services. Yet the dominant engineering practice for such systems remains relatively informal: natural-language prompts, loosely specified tool schemas, and framework-specific orchestration scripts. This opinion article argues that reliable agentic AI requires a transition from prompt engineering to agent engineering: the deliberate design of software abstractions that make agent behaviour interoperable, inspectable, auditable, and governable. I propose a three-layer stack: (i) ~a protocol layer for standardising model--tool, model--data, and agent--agent communication; (ii) ~a workflow and composition layer for encoding state, control flow, evaluation, recovery, and human checkpoints as explicit software artefacts; and (iii) ~a declarative intent layer for specifying goals, constraints, permissions, evidence requirements, and accountability rules without scripting each reasoning step. The article situates emerging developments--including the Model Context Protocol, tool/function calling, Agent2Agent coordination, LangGraph, AutoGen, DSPy, and constitutional alignment--within this broader architectural trajectory. It also identifies open research problems in capability semantics, provenance, security, probabilistic correctness, specification elicitation, compositional assurance, and governance-aware runtime design. The central claim is that the next phase of agentic AI will not be defined only by larger models, but also by the abstractions that make agent behaviour composable, testable, and accountable.