Control-Plane Placement and Memory Substrates in Modern Agentic AI: A Comparative Method for Prompt-Native Systems, Agent Shells, and Durable Runtimes

Abstract AI agents that autonomously write, test, and modify code are now deployed across architectures that share similar surface features but differ substantially in where and how operating control is exercised. Contemporary systems span general-purpose SDKs, prompt-native work-flow packs, interactive coding shells, hosted agents, memory services, and durable workflow runtimes, yet the existing language of framework comparison often treats them as interchangeable. The remaining problem is comparative precision: these systems still look artificially similar if operational control and durable experience are assumed to live in the same architectural layer. This paper proposes control-plane placement, complemented by memory substrate analysis, as a comparative method for modern agentic systems. The organising principle is the enforcement set – the collection of architectural layers that can independently constrain agent behaviour, and the five system families compared here represent its five stable configurations. The method asks where workflow policy, tool exposure, permissioning, routing, verification, and recovery are specified and enforced, and separately how durable experience is stored, searched, updated, and reused across sessions. Using official documentation and open-source implementation evidence, the paper compares prompt-native systems, prompt-modular shells, runtime-governed shells, host-mediated or coordinator forms, and durable workflow runtimes across a twelve-system ecosystem. It contributes a qualitative coding matrix for control-plane salience and a companion memory-oriented comparison covering artefact memory, procedural memory, write-policy visibility, and traceability. The paper also argues that coding-agent benchmarks are best read as evaluations of model × reasoning policy × harness/runtime × memory substrate × environment under explicit tool, budget, and infrastructure conditions. The result is a more discriminating vocabulary for agentic AI and a clearer method for comparing rapidly converging systems.