A new paper on arXiv examines the recent wave of protocols designed to govern how AI agents interact with users, external tools, and one another. Among these, the Model Context Protocol (MCP) has drawn particular attention. The authors argue that this moment in AI development resembles an earlier era in networking, when TCP/IP was emerging as a foundational internet standard.
The paper's core premise is that the security and governance challenges faced by internet protocols offer useful precedents for agentic AI. Rather than treating AI protocols as a wholly new problem, the authors suggest that lessons from TCP/IP's evolution—such as how trust, authentication, and control were handled—can inform the design of safer and more accountable agent systems.
Because the abstract is truncated, the full set of recommendations is not visible in the source. Still, the framing is clear: as MCP and similar protocols mature, their architects would do well to study the successes and failures of the internet's own protocol stack. The paper stops short of prescribing specific fixes, but it makes a case for historical awareness in AI infrastructure design.