The Exploration of AI Agents in the Web3 Field: From Manus to MCP

Recently, a universal AI Agent product called Manus has sparked heated discussions in the tech circle. As the world's first product of its kind, Manus demonstrates a strong ability for independent thinking, planning, and executing complex tasks, capable of autonomously completing the entire process from planning to execution, such as writing reports and creating spreadsheets. The product's explosive popularity has not only attracted industry attention but also provided valuable product ideas and design inspiration for various AI Agent development.

AI Agent is an important branch of artificial intelligence, gradually moving from concept to real-world applications. It is a computer program that can autonomously make decisions and execute tasks based on the environment, inputs, and predefined goals. The core components of AI Agent include large language models (LLM), observation and perception mechanisms, reasoning and thinking processes, action execution, as well as memory and retrieval functions.

The design patterns of AI Agents mainly have two development routes: one focuses on planning capabilities, including REWOO, Plan & Execute, and LLM Compiler; the other emphasizes reflective capabilities, including Basic Reflection, Reflexion, Self Discover, and LATS. Among them, the ReAct pattern is the earliest and most widely used design pattern, with a typical process that includes three steps: thinking, acting, and observing, forming a cyclical process.

According to the number of agents, AI Agents can be divided into Single Agent and Multi Agent. Single Agent focuses on the combination of LLM and tools, while Multi Agent assigns different roles to different agents to complete complex tasks through collaborative cooperation. Currently, most frameworks are concentrated on the Single Agent scenario.

Model Context Protocol (MCP) is an open-source protocol launched by Anthropic, aimed at addressing the connection and interaction issues between LLMs and external data sources. MCP provides three capabilities: knowledge expansion, function execution, and pre-written prompt templates, utilizing a Client-Server architecture with the underlying JSON-RPC protocol.

In the Web3 industry, the development of AI Agents has experienced peaks and declines. Currently, there are three main models: the launch platform model represented by Virtuals Protocol, the DAO model represented by ElizaOS, and the commercial company model represented by Swarms. Among them, the launch platform model is currently the most likely to achieve a self-sustaining economic loop.

The emergence of MCP has brought new exploration directions for Web3's AI Agent. First, deploying the MCP Server to the blockchain network addresses the single point issue and has anti-censorship capabilities; second, it gives the MCP Server the ability to interact with the blockchain, lowering the technical threshold. Additionally, there is a plan to build an OpenMCP.Network creator incentive network based on Ethereum.

Although the integration of MCP with Web3 can theoretically inject decentralized trust mechanisms and economic incentives into AI Agent applications, the current technology still faces challenges. Zero-knowledge proof technology still struggles to verify the authenticity of Agent behavior, and decentralized networks also have efficiency issues.

The integration of AI and Web3 is an inevitable trend. Although there are challenges currently, we must remain patient and confident, continuously exploring the possibilities in this field. In the future, a milestone product may emerge in the Web3 world, breaking the skepticism about the practicality of Web3 from the outside and promoting the application and development of AI Agents in decentralized environments.