From Manus and MCP to AI Agent: Exploring Web3 Cross-Border Innovation

Written by: pignard.eth, ZAN Team

On March 6, a global first general-purpose AI Agent product called Manus, launched by Chinese startup Monica, went viral across China's tech media and social networks. On its first day of release, invitation codes became extremely scarce online—so much so that one code was reportedly selling for 50,000 RMB on Xianyu (a secondhand marketplace). Nonetheless, numerous industry KOLs received early access, triggering a flood of hands-on reviews and analytical articles.

As a general-purpose AI Agent product, Manus can autonomously complete tasks from planning to execution, such as writing reports or creating spreadsheets. It does not merely generate ideas—it thinks independently and takes action. With its powerful capabilities in independent reasoning, task planning, and complex execution, Manus delivers complete results, demonstrating unprecedented versatility and operational strength.

The sudden popularity of Manus has drawn significant attention within the industry and provided valuable product insights and design inspiration for various AI Agent developers. As AI technology advances rapidly, AI Agents—an important branch of artificial intelligence—are gradually transitioning from concept to reality, showing immense application potential across industries, including Web3.

Background Knowledge

An AI Agent, or artificial intelligence agent, is a computer program capable of making autonomous decisions and executing tasks based on environmental inputs and predefined objectives. The core components of an AI Agent include a large language model (LLM) acting as its "brain," enabling it to process information, learn from interactions, make decisions, and perform actions; observation and perception mechanisms allowing it to sense the environment; reasoning processes involving analysis of observations and memory to consider possible actions; action execution as explicit responses to reasoning and observation; and memory and retrieval systems to store past experiences for learning purposes.

AI Agent design patterns evolved from ReAct along two main development paths: one emphasizing planning capabilities, including REWOO, Plan & Execute, and LLM Compiler; the other focusing on reflection abilities, including Basic Reflection, Reflexion, Self Discover, and LATS.

ReAct is the earliest and currently most widely used AI Agent design pattern, so we will focus on explaining ReAct here. ReAct refers to solving diverse language reasoning and decision-making tasks by combining reasoning and acting within language models. Its typical workflow can be described through an interesting loop: Thought → Action → Observation, commonly known as the TAO cycle.

• Thought: When facing a problem, deep thinking is required—defining the problem, identifying key information, and determining reasoning steps necessary for resolution.

• Action: After establishing the direction of thought, corresponding measures are taken or specific tasks executed to advance toward solving the problem.

• Observation: Following action, careful observation of outcomes is essential to evaluate whether the action was effective and brought us closer to a solution.

• Iterative Looping

AI Agents can also be categorized by the number of agents involved: Single Agent and Multi Agent. In Single Agent systems, the core lies in the coordination between the LLM and tools, often involving multiple rounds of interaction with users during task completion. In Multi Agent setups, different agents are assigned distinct roles and collaborate to accomplish complex tasks, typically requiring less direct user interaction than Single Agent systems. Most current frameworks focus primarily on Single Agent scenarios.

Model Context Protocol (MCP), introduced by Anthropic on November 25, 2024, is an open-source protocol designed to address connectivity and interaction challenges between LLMs and external data sources. Think of the LLM as an operating system and MCP as a USB interface—supporting flexible plug-in of external data and tools, which users can then read and utilize.

MCP extends LLM capabilities through three functions: Resources (knowledge expansion), Tools (execution functions for calling external systems), and Prompts (pre-written prompt templates). The MCP protocol adopts a Client-Server architecture with JSON-RPC as the underlying transport protocol. Anyone can develop and host an MCP Server and may take it offline at any time.

Status of AI Agents in Web3

In the Web3 space, enthusiasm around AI Agents peaked in January this year but has since declined sharply, with overall market capitalization shrinking by over 90%. Currently, the most prominent projects exploring AI Agents in Web3 fall into three models: “launchpad platforms” represented by Virtuals Protocol, “DAO models” exemplified by ElizaOS, and “commercial company models” like Swarms.

A launchpad platform enables users to create, deploy, and monetize AI Agents—similar to pump.fun in the Meme coin space, but tailored for AI Agents. Virtuals Protocol is currently the largest such platform, hosting over 100,000 Agents. One particularly popular Agent, AIXBT—a crypto KOL persona—is built on Virtuals. The platform features a modular Agent framework called G.A.M.E., positioned as an efficient and open framework that makes developing and launching AI Agents as simple as building websites with WordPress.

DAO stands for Decentralized Autonomous Organization. ElizaOS (formerly ai16z), founded by @shawmakesmagic on daos.fun, initially aimed to simulate investment decisions of renowned venture capital firm a16z and its co-founder Marc Andreessen using AI models, combined with suggestions from DAO members. It later evolved into a developer-focused DAO centered around the Eliza framework. Built with TypeScript, the Eliza framework provides a flexible and scalable platform for developing AI Agents that maintain consistent personality and knowledge while interacting across multiple platforms.

Swarms, initiated in 2022 by 20-year-old @KyeGomezB, is an enterprise-grade Multi Agent framework. Through intelligent orchestration and efficient collaboration, Swarms enables multiple AI Agents to work together like a team, addressing complex business operations. Initially a Web2 AI Agent project, Swarms claims to have over 45 million agents running in production environments, serving some of the world’s largest financial, insurance, and healthcare institutions. It officially transitioned to Web3 after launching its token $SWARMS in December 2024.

From an economic model perspective, only the launchpad platforms currently achieve self-sustaining economic loops. Taking Virtuals as an example:

1. Agent Creation: Creators initiate new AI Agents on the Virtuals platform;

2. Bonding Curve Setup: Creators pay 100 $VIRTUAL tokens to establish a bonding curve for the new agent’s token, paired with $VIRTUAL.

3. Liquidity Pool Creation: Once the bonding curve limit is reached, the agent "graduates," and a liquidity pool is created pairing the agent’s token with $VIRTUAL, adhering to fair launch principles—no pre-mine or insider allocation, fixed total supply, and long-term locked liquidity.

Besides charging fees for launching AI Agents, Virtuals also collects transaction fees on every trade of agent tokens and inference fees when Agents access LLMs via Virtuals’ API. Both ElizaOS and Swarms are now planning to build their own launchpad platforms.

However, launchpad platforms face issues too. This asset issuance model relies on the issued assets being inherently “attractive” to create a positive flywheel effect. Most launched AI Agents today are essentially Memes without intrinsic value. Once they lose market attention, their value quickly collapses to zero. In the current bleak market conditions, even attracting creators to these platforms has become difficult, rendering the economic models practically non-functional.

Web3 Explorations of MCP

The emergence of MCP opens new exploration avenues for Web3-based AI Agents, mainly in two directions:

1. Deploying MCP Servers onto blockchain networks to solve single-point failure issues while achieving censorship resistance;

2. Equipping MCP Servers with blockchain interaction capabilities—for instance, enabling DeFi trading and management—to lower technical barriers.

The first direction demands extremely high requirements on the underlying blockchain’s storage system, data management, and asynchronous computing capabilities. Blockchains like 0G could be suitable choices. 0G is a modular AI blockchain featuring a scalable and programmable data availability (DA) layer optimized for AI dApps. Its modular architecture enables frictionless interoperability between chains while ensuring security, eliminating fragmentation, maximizing connectivity, and fostering a decentralized AI ecosystem.

The second direction resembles variations of DeFAI, though current DeFAI backends rely on custom-wrapped Function Call tools. UnifAI aims to create a unified DeFAI MCP Server, avoiding redundant development. UnifAI is a platform enabling autonomous AI Agents to perform both on-chain and off-chain tasks within the Web3 ecosystem. It includes UniQ for task automation, an agent service marketplace, and infrastructure for tool discovery.

Beyond these two directions, @brucexu_eth, founder of LXDAO and ETHPanda, proposed a concept called OpenMCP.Network—an incentive network for creators built on Ethereum. Since MCP Servers require hosting and stable service provision, users would pay LLM providers who, in turn, distribute actual incentives through the network to the invoked MCP Servers, sustaining the network’s stability and sustainability and motivating creators to continuously produce high-quality content. This network would use smart contracts to automate incentives in a transparent, trustworthy, and censorship-resistant manner. Signatures, permission verification, and privacy protection during operation could be implemented using Ethereum wallets, ZK technologies, and more.

Theoretically, integrating MCP with Web3 can inject decentralization-based trust mechanisms and economic incentive layers into AI Agent applications. However, current zero-knowledge proof (ZKP) technology struggles to verify the authenticity of Agent behaviors, and decentralized networks still face efficiency challenges—making this a long-term rather than short-term viable solution.

Conclusion

The release of Manus marks a significant milestone in the development of general-purpose AI Agent products. Similarly, the Web3 world needs a landmark product to counter external skepticism about Web3 lacking real utility and being driven purely by hype.

The emergence of MCP brings fresh directions for Web3 AI Agents—whether deploying MCP Servers on blockchains, enabling them to interact with blockchains, or building creator incentive networks for MCP Servers.

AI represents the grandest narrative in history. For Web3, convergence with AI is inevitable. We must remain patient and confident, continuing our explorations forward.