The Internet of Agents is not equal to multi-agent collaboration: Analysis of the network effects of MCP, ANP and A2A

Written by

Caleb Hayes

Updated on:July-01st-2025

1 Agents as Network Nodes Some people compare agents to websites on the Internet. It is true that in the future, each agent can register a unique identity and provide accessibility. However, agents are no longer passive information pages waiting to be accessed, but rather actors who are good at taking the initiative, calling tools, and helping the client to complete his or her intentions. Websites are statically connected through hyperlinks, while agents are dynamically connected through actions. This is the fundamental difference between the Agentic Web and the Web. How will this dynamic connection happen? One scenario is that changes in consumer behavior will drive supply-side changes. For example, personal financial assistants help consumers send emails to service operators requesting discounts. In response, operators can deploy customer service agents to automatically reply to various letters based on their internal knowledge base. Another scenario is that the intelligentization of the supply side drives changes in the consumer side. For example, if each e-commerce website provides a virtual shopping assistant, it will be very efficient to use a consumer assistant to interact with these different shopping assistants. Of course, if all brands deploy e-commerce agents, it will also stimulate the popularization of consumer-side agents, and even no longer require centralized e-commerce websites. This direct interaction between consumer-side and supply-side agents is the main form of dark traffic in the future. Network effects emerge: more and more consumer-side agents will prompt the release of supply-side agents; more and more supply-side agents will stimulate innovation and diversity of consumer-side agents. 2 There are two network effects in the Internet of Agents. The first is the network effect between agents and tools connected by MCP. The more agents there are, the more MCP tools there are, because tool providers find the 2A market profitable; the more MCP tools there are, the more agents there are, because more agent developers' creativity can be supported by tools. As a result, agents become more and more powerful. For example, compared with a travel planning assistant that only calls a search tool, calling a map can give a more reliable and accurate journey plan, and calling the ticket information of TravelSky can further give round-trip air ticket recommendations. The second is the network effect between agents. Similar to the email interaction between consumers and operators mentioned above, TravelSky's ticket sales agent can trade air tickets with the above travel planning assistant. As a result, the agent network becomes more and more extensive, carrying more and more economic activities, and the agent economy takes shape. The open source protocol that supports this network effect is ANP, the Agent Network Protocol. The difference between these two types of network effects is similar to that between one-way and two-way networks. MCP supports one-way networks centered on models, that is, models can call tools, but tools cannot actively communicate with models. ANP supports P2P interconnection between agents, and any agents can communicate with each other. In the above example, the difference between the MCP tool provided by TravelSky and the ticket sales agent provided by TravelSky is that the tool can only be called passively, while the ticket sales agent can actively contact the travel assistant agent. For example, the operator's customer service agent can actively contact the personal financial assistant agent to recommend the latest packages and discounts. Precisely because they need to support different interaction modes, there are some key technical differences between ANP and MCP. 3 Multi-agent collaboration and agent Internet Recently, Google released the A2A (Agent 2 Agent) protocol, which triggered a wave of self-media explosion. As far as A2A is a communication protocol for agents, the above analysis of ANP and MCP is fully applicable to understanding the difference between A2A and MCP.

But what is the difference between ANP and A2A? I think A2A is better at supporting multi-agent collaboration, while ANP is better at supporting the Internet of Agents.

Multi-agent collaboration is a tight coupling towards a common complex task, that is, the division of labor of each agent in this task network usually has a relatively close interdependence. A2A has an advantage in this regard because it maintains real-time dynamic synchronization of task status among multiple agents, thereby ensuring the transparency of the collaboration process. The sponsor behind A2A is Google Cloud, and its supporters are mainly 2B service providers, including SaaS providers, IT service providers, and consulting companies. Not surprisingly, these companies pay more attention to multi-agent collaboration within the enterprise. But the Internet of Agents is a loosely coupled ecosystem without boundaries and predefined tasks. In terms of supporting the Internet of Agents, ANP provides a more native identity registration and authentication method, supporting any agent to create an independent identity and interoperability between any two agents. In contrast, A2A itself does not provide an identity registration mechanism, but allows each agent to reuse the existing authentication method and declare it in the Agent Card. For these technical details, the analysis of ANP author Chang Gaowei is more professional than mine. If you are interested, you can refer to this post . I want to emphasize that multi-agent collaboration and the Internet of Agents are different concepts, although technically both involve communication between agents. Multi-agent collaboration is similar to the division of labor between people - any economy cannot do without division of labor, and division of labor will promote growth, but as emphasized in the previous analysis of this article, the core economic driving force of the Internet of Intelligent Agents comes from network effects - division of labor does not necessarily bring about network effects. If all divisions of labor have network effects, we would not have to wait until the rise of the platform economy in the 21st century to realize the huge charm of network effects. I have written these initially. After Chang Gaowei comes to the AI Collision Bureau of the National Development Institute on Tuesday night, I will update, revise and expand. Of course, everyone is also welcome to come to the National Development Institute to participate in the collision.