Agentic AI Architectural Principles for Autonomous Computer Networks

Agentic AI systems combine planning, reasoning, tool invocation, and feedback loops to pursue system-defined goals with a controlled degree of autonomy. In networking, this enables an evolution from statically configured automation toward goal-driven closed-loop operations spanning multiple protocol layers and administrative domains. This document introduces architectural principles for "agentic augmentation" of the existing layered protocol stack as represented by the Internet protocol suite (IP suite). The key concept of the proposed principles is that deterministic protocol layering remains intact for interoperability, while AI Agents are introduced as first- class entities at each IP suite layer and are coordinated by one or more agent controllers via agentic methods and procedures. The purpose of this document is to initiate discussion within the research community on agentic networking. It identifies architectural research challenges that should be discussed to enable the addition of one or more AI Agents at one or more IP suite layers with the goal to allow AI Agents to improve the behaviour of a layer through reasoning with AI Agents at the same or other IP suite layers.