Agents have essentially become the consensus of the AI industry, and if there are differences, they are most likely regarding the timeline for implementation.
The current focus of the industry is still on how to build agents themselves, and there is not much research on how agents can collaborate or the characteristics of the agent network.
In this article, we attempt to explore the differences between the Agentic Web and the existing internet from the essence of internet evolution: the ability to unleash new technologies. The focus of this discussion is on the connectivity and collaboration of the internet, and we will not elaborate on other aspects.
Whether the Agentic Web will emerge or whether it will be called that, we are uncertain and not overly concerned. What we are very certain about is that agents will bring profound changes to the internet.
Given the current global internet market size, the differences that the Agentic Web presents is at least a hundred billion dollar question.
Key points of this article:
-
The most fundamental question regarding the direction of internet evolution is how to maximize the ability to unleash new technologies. The next generation of the internet will undoubtedly undergo profound changes due to the development of AI technologies. -
To unleash the capabilities of AI, we need to achieve the following four points: 1) AI must obtain complete contextual information; 2) AI must be able to access all tools on the internet; 3) AI must be able to interact with the internet in the way it is most proficient; 4) AI must be able to collaborate efficiently with other AIs on the internet. -
This requires that the future internet must be an open and interconnected internet, breaking the current situation of data silos. -
The way AI interacts with the internet most effectively is through APIs or communication protocols. It is better at handling underlying data rather than graphical interfaces. -
We are not optimistic about Computer Use technology because it essentially makes AI mimic humans, forcing AI to conform to the current state of the internet. Of course, we do not deny that this has commercial value in the short to medium term. -
AI smartphones provide a solution to connect various apps at the terminal, but the relationship between AI smartphones and apps is somewhat awkward. Most future applications will not require a terminal interface; personal assistants will provide personalized UIs for everyone. -
Anthropic’s MCP is the technology that most aligns with our vision and is a direction we are very optimistic about. We have been working on similar projects for a long time. -
Future vision: Personal assistants will replace humans in accessing the internet, becoming the new entry point; agents will permeate every corner of the internet; all agents can interconnect and communicate, building an efficient collaborative network; the internet will become more decentralized and flattened.
What is the fundamental issue determining the direction of internet evolution?
We believe that the most fundamental issue driving the evolution of the internet, or the underlying driving force, is how to unleash the capabilities of new technologies.
Analyzing a question, why were products during the PC internet era primarily web-based, while mobile internet era products are primarily app-based?
During the PC internet era, new technologies were primarily based on network communication technologies (TCP/IP, DNS, HTTP), web technologies (HTML, CSS, JavaScript), database technologies, information search, and so on. Products mediated by browsers and web technologies unleashed the access capabilities of the open internet, allowing information to be widely obtained through a unified entry point.
Moreover, the technological conditions at that time (limited bandwidth and hardware performance) and the cross-platform characteristics of web technologies determined that the web was the most suitable product form for the PC internet.
By the time of the mobile internet, new technologies included mobile devices and hardware (accelerometers, gyroscopes, GPS, touch screens, cameras, chips), mobile networks (3G, 4G, 5G), and mobile operating systems (iOS, Android), among others.
Apps, compared to web technologies, can deeply integrate hardware capabilities through operating system APIs, optimize operating efficiency for device performance, and remain online continuously while interacting in real-time.
It is these factors that allowed new technological capabilities to be unleashed through apps, and the app experience far exceeded that of the web, making apps the mainstream product form of the mobile internet era.
In the present and the next decade, new technologies are undoubtedly represented by generative artificial intelligence, large language models, and agents. How to unleash the capabilities of AI will determine the direction of future internet evolution.
How should the next generation of the internet unleash AI’s capabilities?
The current internet infrastructure is already quite mature, but to fully unleash AI’s capabilities, there are still some challenges to face.
How to enable AI to obtain complete contextual information and how to allow AI to access all tool capabilities
AI can only make correct decisions if it obtains complete contextual information. Only by being able to access all tool capabilities can it efficiently complete complex tasks.
However, the current internet is essentially composed of isolated information islands, making it difficult for information to flow between these islands.
Previously, humans actually played the role of a patchwork for information islands, connecting these islands through browsers, apps, search engines, social networks, etc. In the future, this will be accomplished more efficiently by AI.
There are already technical solutions attempting to solve these issues, such as AI smartphones and Computer Use technology, allowing AI to connect with the internet through graphical interfaces, browsers, or app terminal interfaces. However, we believe these are not the most efficient solutions.
How to allow AI to access the internet in the way it is most proficient?
Let us comment on the current solutions for AI accessing the internet.
-
Computer Use technology
Many model vendors have launched similar solutions.
However, I have never been optimistic about this technology. This technology emerged because existing internet products were designed for human use. Before these products are restructured for AI, having AI learn and mimic humans is indeed the fastest way for AI to access the internet. However, it is not the most efficient path. It may have value in the short to medium term, but not in the long term.
-
AI smartphone solutions
For example, Apple Intelligence, many domestic smartphone manufacturers are also vigorously promoting.
The characteristic of this type of technology is that, at the terminal, it opens interfaces through APP clients to connect the internet’s data and capabilities. This allows AI to obtain data from multiple apps and utilize the capabilities of multiple apps.
This technology is a step further than Computer Use technology; AI is inherently better at processing underlying data rather than graphical interfaces. However, in this solution, the positioning of apps on the terminal is quite awkward, as they are in a competitive yet cooperative relationship with AI smartphones, wanting to gain the traffic brought by AI smartphones while not wanting to open too much data to be replaced by AI smartphones.
In the long run, many apps will not need a user interface at the terminal; personal assistants will provide personalized UIs for everyone.
-
Anthropic MCP
This is currently the solution that best aligns with our vision.
We believe that AI is different from humans; it is better at processing underlying data rather than graphical interfaces. AI should interact with the internet in the way it is most proficient (APIs or communication protocols).
We have also been working on similar projects and research for a long time: Agent Network Protocol technical white paper.
The biggest difference between us and MCP lies in the difference in worldview:
-
MCP is model-centric, and the entire internet serves as its context and tools; -
We (Agent Network Protocol) are agent-centric, where each agent has equal status, forming a decentralized agent collaboration network.
How to allow AI to collaborate efficiently with other AIs on the internet
Traditional network nodes often connect through hard coding or manual methods. With AI, the connection and collaboration methods of the network may become more efficient.
For example, two AI nodes can utilize their natural language generation and understanding capabilities to first communicate their abilities and interfaces in natural language, and then use standard protocols or consensus protocols for communication and collaboration.
This will help build a more efficient and lower-cost collaborative network.
To summarize, if the internet wants to fully unleash AI’s capabilities, it still needs to solve the following three problems:
-
🌐Interconnectivity: Allow all agents to communicate with each other, breaking data silos, enabling AI to obtain complete contextual information and access all tools on the internet. -
🖥️Native interfaces: AI should not imitate human access to the internet; it should interact with the internet in its most proficient way (APIs or communication protocols). -
🤝Efficient collaboration: Utilizing AI, agents can self-organize and self-negotiate, building a collaboration network that is lower cost and more efficient than the current internet.
What is the Agentic Web?
The definition of the Agentic Web is a rather large proposition, and we cannot provide a rigorously academic definition. However, we can explain our understanding from a common perspective.
-
Agents become a new entry point
The PC internet is humans using PCs to access the internet; the mobile internet is humans using mobile devices to access the internet; the Agentic Web, however, is agents replacing humans in accessing the internet and interacting with the digital world.
This entry point will likely be a special agent in the future: a super personal assistant. It will replace humans in accessing the internet and interact with humans through a personalized UI at the client side while interacting with other agents via APIs or protocols at the backend.
-
Agents permeate every corner of the internet
In addition to personal assistants, we believe that there are many agents on the internet that do not directly interact with end-users, such as those in hotels, restaurants, banks, and schools. They provide services indirectly to people through interactions with personal assistants.
-
Agents can interconnect
Whether personal assistants or the agents providing services behind the scenes, regardless of which company or platform they belong to, they can interconnect.
-
A flatter network
If agents can connect and collaborate as nodes in the internet on behalf of a person or entity, the role of centralized platforms (like WeChat, Taobao) will diminish. The future internet will be more decentralized and flattened.
-
A self-organizing and self-negotiating efficient collaborative network
Agents can self-organize and self-negotiate to build an efficient collaborative network.
Conclusion
Ultimately, we believe that the development of technology drives the evolution of the internet, and how to unleash the capabilities of new technologies determines the direction of internet evolution.
Whether the Agentic Web will emerge is still uncertain, but we can be sure that the future internet will undergo profound changes due to the development of AI technologies.
Contact Us
We are currently developing an open standard protocol for agent communication based on our vision: AgentNetworkProtocol. Project address: https://github.com/chgaowei/AgentNetworkProtocol.
The goal of AgentNetworkProtocol is to become the HTTP of the agent internet era. Our vision is to define the connection methods between agents, building an open, secure, and efficient collaborative network for billions of agents.
If you are also interested in agent communication protocols or the agent internet, please feel free to contact us. We are currently sharing and exchanging ideas with AI practitioners domestically and internationally:
-
WeChat: flow10240 -
Jike: 常高伟_长山 -
Email: [email protected]
In conclusion, I believe that if scalability is not considered, web3 may be the most suitable internet form for AI. If you are working on web3 + AI, we welcome you to exchange ideas with us.
Related Articles
Technical White Paper on Agent Network Protocol: A Protocol Benchmarking Anthropic MCP
Anthropic MCP: A Technology Overestimated in the Short Term, Underestimated in the Long Term
My Understanding of AI-Native Applications and New Platform Paradigms
The Most Efficient Way to Consume Information in the Era of Large Models