Agent's Impact on Infrastructure: Challenges to Existing Connection Infrastructure
AI's Special Requirements for Connectivity
For intelligent agents built on LLMs to maximize their capabilities, they have special requirements for connectivity:
- AI must be able to access all information to make more intelligent decisions for people or organizations, requiring AI to be interconnected with all information.
- AI must be able to call all tools to better help people or organizations complete tasks, requiring AI to be interconnected with all tools.
Currently, all apps, software, or websites are designed for human access. For AI to access them, it must learn and mimic human access methods. This has led to the emergence of Computer use and Browser use technologies.
These two technologies are valuable in the short term, but they have inherent limitations. For example, Computer use is costly due to visual information based on screens and requires occupying the user's computer screen; Browser use can only access websites and has security issues during user login.
Therefore, we need more AI-native ways for AI to interact with the digital world and for intelligent agents to interact with each other.
Protocol: AI-Native Connection and Collaboration Method
This has given rise to another technological approach: Protocol.
Using Protocol to process data is more efficient: AI is better at handling the most direct underlying data, rather than screens or browsers.
Standardization can reduce connection costs: By standardizing Protocol, all software or intelligent agents that support standard Protocol can interconnect, greatly reducing connection costs and forming a new ecosystem.
Using protocols, an AI-native data network that is easily accessible to AI can be built, independent of the current internet designed specifically for humans, based on screens and browsers.
I believe this is the underlying reason why MCP has recently gained widespread attention. It solves the most important current problem in an AI-native way: how models connect to tools and resources. At the same time, standardization can reduce the connection costs of the entire industry, forming a new ecosystem.
After OpenAI announced support for MCP, MCP can basically be seen as the industry's de facto standard, and it's unlikely that a second MCP will emerge. Future cooperation between MCP and standardization organizations to promote MCP standardization is likely.
The biggest problem with the Protocol approach is that it requires modifying existing software, which is costly. However, this cost is being rapidly reduced by AI:
- AI is changing the way software is developed; through AI-generated code, software development costs can be significantly reduced.
- AI is changing the way software runs; using LLM to drive software operation can solve the essential complexity of software within the LLM, eliminating the need for large amounts of complex code.
- AI is changing the way software connects; by using LLM to construct requests and process responses, the cost of software integration and joint debugging can be greatly reduced.
MCP's Next Step: Agent Communication and Collaboration
Using MCP can build more efficient agents, but as agents become more numerous, how do they communicate and collaborate?
MCP is designed to solve the problem of connecting models to resources and tools, and is difficult to use directly for communication and collaboration between agents. There are two core problems (taking agent A accessing agent B as an example):
- Identity problem: Using MCP requires agent A to have a registered account with agent B. If agent A needs to access many agents belonging to different organizations, A must manage many accounts, which is very costly.
- Protocol architecture: MCP is a typical client-server architecture, where the client actively connects to the server, and the roles of client and server are strictly distinguished. In an agent network, the relationship between agents is more peer-to-peer. The client-server architecture is inherently unsuitable for peer-to-peer communication between agents, especially in agent connection and collaboration: one-way client-server connection means one party cannot find the other; two-way client-server connection can solve this problem, but is overall very complex.
There are many protocols in the industry specifically designed to solve the challenges of communication and collaboration between agents. Currently, the most complete in design and implementation is ANP (Agent Network Protocol), as well as IBM's ACP (Agent Communication Protocol) and Cisco's ACP (Agent Connection Protocol).
Taking ANP as an example, it focuses on solving the problem of agent collaboration on the internet, such as agent identity, agent description, and agent discovery. Its biggest feature is using W3C DID (decentralized identity) to identify agents, allowing agents to use their own identity IDs to interact with any other agents. At the same time, ANP's protocol architecture is P2P, naturally suitable for peer-to-peer communication between agents.
The Internet's Next Step: The Internet of Agents
As more and more agents appear on the internet and connections between agents increase, this will inevitably cause profound changes to the existing internet. We have several judgments about the evolution of the internet:
- Personal assistants become the new internet entry point, with people accessing the internet through personal assistants.
- Enterprises will also use agents to replace enterprise software and deploy agents directly on the internet to serve consumers.
- Direct connection between personal assistants and enterprise agents through protocols becomes possible, rather than connecting through internet platforms. Consumer internet will deeply integrate with industrial internet.
- Connections between agents will become the main way of internet connection, and a specialized agent data network designed for AI and easy for AI to access will emerge.
- Agent communication protocols become a new infrastructure, and standardized agent communication protocols will emerge in the future.