Guosheng Securities: MCP and A2A help AI agents enter the connected era

The Zhitong Finance App learned that Guosheng Securities published a research report saying that the market believes that due to limited training data, the pre-training scaling law has hit a wall, and demand for computing power is under pressure. The bank believes that agents will open up the ceiling for inference computing power. Various agencies are actively developing agents, and the competitive landscape is not stable; agent computing power requirements are not the same as simple text responses. Recommended attention: computing power chip vendors, underlying protocol developers, computing power cloud service vendors, and large model vendors.

Guosheng Securities's main views are as follows:

What is an Agent?

Recently, topics related to the agent field, such as Microsoft's launch of the Github MCP server, Google's release of the A2A inter-agent communication protocol, and Alipay's launch of the MCP server, have received widespread attention from the market. There is no uniform definition of an agent. Lilian Weng, a former OpenAI researcher, was well recognized. She suggested that “planning,” “memory,” and “tool use” are important components of agents.

Agent development status: Currently, there are only a few independent charges. The penetration rate is low, but there is more space

As an evolved version of Chat, currently most agents are only included in the larger model of paid service content, with only a few independent charges such as Manus and Devin. Agents such as Deep Research and Manus, which have the ability to plan tasks independently, have many usage restrictions. There are probably not many users who can actually experience it, and there is still plenty of room for improvement from “popular” applications. However, as the big model's reasoning ability improved, agents became the darling of applied innovation.

Large-scale application of agents is imminent, and many favorable conditions are at hand

1) The context window on the model training side is growing rapidly, reinforcement learning is being applied in depth, and inference models are becoming more mature; 2) Ecology: protocols such as MCP and A2A are developing rapidly, and agent calling tools are becoming more and more convenient. In November 2024, Anthropic released and open-sourced the MCP protocol, which aims to standardize how external data and tools provide context for models.

The MCP protocol allows agents to “connect with external data and tools” “with one click”, and A2A interconnects agents

Although the goal of the A2A protocol is communication between agents, and MCP is an agent and external tools and data, the two functions may overlap in complex situations where “tools may also be encapsulated as agents,” but this competition helps reduce the cost of calling external tools and communication for large models.

Agent development outlook 1: End-to-end, no need for humans to write a workflow in advance

Currently, quite a few of the “smart devices” seen by the bank are developed based on platforms such as Coze and Dify, and require humans to write a workflow in advance. They are more elementary agents, more like superposition of prompt engineering, while more advanced agents are “end-to-end”, which means “inputting tasks to agents, agents automatically complete the task results required by humans.” Advanced agents such as L3/L4/L5 are more in line with human needs.

Agent Development Outlook 2: Empowering Robots and Autonomous Driving

When applying the definition of agent to physical intelligence, you will find that robots and vehicles controlled by large models are also agents. In particular, the former, the bottleneck in robot development is not the “cerebellum” of “how to perform physical actions,” but rather the “brain” that thinks “what kind of physical action to perform,” and this is falling within the range of an agent.

Agent development outlook 3: Using DID and other technologies to achieve interconnection between agents and native AI networks

In the future, perhaps all agents should be able to communicate with each other, self-organize, and self-negotiate to build a lower cost and more efficient collaborative network than the existing Internet. The Chinese developer community is also building protocols such as ANP, with the aim of becoming the HTTP protocol for the Agent Internet era. However, for identity authentication between agents, technology such as DID can be used.