AI Agent持续阴跌是由最近爆火的MCP协议造成的?

Odaily星球日报Published on 2025-03-17Last updated on 2025-03-17

Abstract

Manus + MCP才是web3 AI Agent此番遭受冲击的关键。

原文作者:Haotian(X:@tme l0 211 

有朋友说,诸如 #ai16z、 $arc 等 web3 AI Agent 标的的持续阴跌是由最近爆火的 MCP 协议造成的?乍一听,整个人有点懵,WTF 有关系吗?但细想之后发现,真有一定的逻辑:已有 web3 AI Agent 的估值定价逻辑变了,叙事方向和产品落地路线亟需调整!以下,谈谈个人观点:

1)MCP(Model Context Protocol)是一个旨在让各类 AI LLM/Agent 无缝连接到各种数据源和工具的开源标准化协议,相当于一个即插即拔 USB「通用」接口,取代了过去要端到端「特定」封装方式。

简单而言,原本 AI 应用之间都有明显的数据孤岛,Agent/LLM 之间要实现互通有无则需要各自开发相应的调用 API 接口,操作流程复杂不说,还缺乏双向交互功能,通常都有相对有限的模型访问和权限限制。

MCP 的出现等于提供了一个统一的框架,让 AI 应用可以摆脱过去的数据孤岛状态,实现「动态」访问外部的数据与工具的可能性,可以显著降低开发复杂性和集成效率,尤其在在自动化任务执行、实时数据查询以及跨平台协作等方面有明显助推作用。

说到此,很多人立马想到了,如果用多 Agent 协作创新的 Manus 集成此能促进多 Agent 协作的 MCP 开源框架,是不是就无敌了?

没错,Manus + MCP 才是 web3 AI Agent 此番遭受冲击的关键。

2)但,匪夷所思的是,无论 Manus 还是 MCP 都是面向 web2 LLM/Agent 的框架和协议标准,其解决的都是中心化服务器之间的数据交互和协作的问题,其权限和访问控制还依赖各个服务器节点的「主动」开放,换句话来说,它只是一种开源工具属性。

按理说,它和 web3 AI Agent 追求的「分布式服务器、分布式协作、分布式激励」等等中心思想完全背离,中心化的意大利炮怎么能炸掉去中心化的碉堡呢?

究其原因在于,第一阶段的 web3 AI Agent 太过于「web2 化」了,一方面源于不少团队都来自 web2 背景,对 web3 Native 的原生需求缺乏充分的理解,比如,ElizaOS 框架最初就是一个,帮助开发者快捷部署 AI Agent 应用的封装框架,恰恰就是集成了 Twitter、Discord 等平台和一些 OpenAI、Claude、DeepSeek 等「API 接口」,适当封装了一些 Memory、Charater 通用框架,帮助开发者快速开发落定 AI Agent 应用。但较真的话,这套服务框架和 web2 的开源工具有何区别呢?又有什么差异化优势呢?

呃,难道优势就是有一套 Tokenomics 激励方式?然后用一套 web2 可以完全取代的框架,激励一批更多为了发新币而存在的 AI Agent?可怕。。顺着这个逻辑看,你就大概明白,为何 Manus +MCP 能够对 web3 AI Agent 产生冲击?

由于一众 web3 AI Agent 框架和服务只解决了类同 web2 AI Agent 的快捷开发和应用需求,但在技术服务和标准和差异化优势上又跟不上 web2 的创新速度,所以市场 / 资本对上一批的 web3 的 AI Agent 进行了重新估值和定价。

3)说到此,大致的问题想必找到症结所在了,但又该如何破局呢?就一条路:专注于做 web3 原生的解决方案,因为分布式系统的运转和激励架构才是属于 web3 绝对差异化的优势。

以分布式云算力、数据、算法等服务平台为例,表面上看似这种以闲置资源为由头聚合起来的算力和数据,短期根本无法满足工程化实现创新的需要,但在大量 AI LLM 正在拼集中化算力搞性能突破军备竞赛的时候,一个以「闲置资源、低成本」为噱头的服务模式自然会让 web2 的开发者和 VC 天团不屑一顾。

但等 web2 AI Agent 过了拼性能创新的阶段,就势必会追求垂直应用场景拓展和细分微调模型优化等方向,那个时候才会真正显现 web3 AI 资源服务的优势。

事实上,当以资源垄断方式爬上巨头位置上的 web2 AI 到一定阶段,很难再退回来用农村包围城市的思想,逐个细分场景击破,那个时候就是过剩 web2 AI 开发者 + web3 AI 资源抱团发力的时候。

事实上,web3 AI Agent 除了 web2 的那套快捷部署 + 多 Agent 协作通信框架外 +Tokenomic 发币叙事之外,有很多 web3 Native 的创新方向值得去探索:

比如,配备一套分布式共识协作框架,考虑到 LLM 大模型链下计算 + 链上状态存储的特性,需要诸多适配性的组件。

1、一套去中心化的 DID 身份验证系统,让 Agent 能够拥有可验证的链上身份,这像执行虚拟机为智能合约生成的唯一性地址一样,主要为了后续状态的持续追踪和记录;

2、一套去中心化的 Oracle 预言机系统,主要负责链下数据的可信获取和验证,和以往 Oracle 不同的是,这套适配 AI Agent 的预言机可能还需要做包括数据采集层、决策共识层、执行反馈层多个 Agent 的组合架构,以便于 Agent 的链上所需数据和链下计算和决策能够实时触达;

3、一套去中心化的存储 DA 系统,由于 AI Agent 运行时的知识库状态存在不确定性,且推理过程也较为临时性,需要一套把 LLM 背后的关键状态库和推理路径记录下来存储于分布式存储系统中,并提供成本可控的数据证明机制,以确保公链验证时的数据可用性;

4、一套零知识证明 ZKP 隐私计算层,可以联动包括 TEE 时、FHE 等在内的隐私计算解决方案,实现实时的隐私计算 + 数据证明验证,让 Agent 可以有更广泛的垂直数据来源(医疗、金融),继而 on top 之上有更多专业定制化的服务 Agent 出现;

5、一套跨链互操作性协议,有点类似于 MCP 开源协议定义的框架,区别在于这套 Interoperability 解决方案,需要有适配 Agent 运行、传递、验证的 relay 和通信调度机制,能够完成 Agent 在不同链间的资产转移和状态同步问题,尤其是包含 Agent 上下文和 Prompt、知识库、Memory 等复杂的状态等等;

在我看来,真正的 web3 AI Agent 的攻克重点应该在于如何让 AI Agent 的「复杂工作流」和区块链的「信任验证流」如何尽可能契合。至于这些增量解决方案,由已有的老叙事项目升级迭代而来,还是由新构成的 AI Agent 叙事赛道上的项目重新铸就,都有可能性。

这才是 web3 AI Agent 应该努力 Build 的方向,才是符合 AI +Crypto 大宏观叙事下的创新生态基本面。若不能有相关的创新开拓和差异化竞争壁垒建立,那么,每一次 web2 AI 赛道的风吹草动,都可能搅得 web3 AI 天翻地覆。

原文链接

Trending Cryptos

Related Reads

Doubao and Qwen Will Discontinue Agent Functionality on July 15

On July 4th, Doubao and Tongyi Qianwen announced the impending shutdown of their user-created "AI Agent" features. Doubao confirmed its agent feature will be taken offline on July 15, directing users to ByteDance's CatBox app for similar needs. On the same day, Tongyi Qianwen notified users, specifying that personalized interactive agents and user-built agent functions will cease on July 10, with all agent features and services completely deactivated by July 15. After this date, access to agent configurations and historical chat records will be lost. This adjustment impacts core user scenarios like role-playing, personal assistants, and vertical tool agents. The shutdown date coincides with the official implementation of China's "Interim Measures for the Administration of Artificial Intelligence Human-like Interactive Services" on July 15. The new regulations impose strict rules on "human-like emotional interaction services," requiring platforms to implement measures like anti-addiction systems, minor verification, and content moderation. This move is widely seen as a proactive step by the platforms to align with regulatory timelines and mitigate compliance risks. Additionally, commercial challenges are a key driver. Analysis suggests that casual, human-like chat agents generate high-frequency, low-value interactions, leading to high computational costs with poor monetization. As the AI application market shifts from user growth to proving value, sustaining such "high-cost, low-efficiency" user-generated content becomes difficult. Both platforms have outlined transition plans. Doubao will allow data viewing and self-backup for a period after shutdown, with data scheduled for permanent deletion by October 15. Tongyi Qianwen similarly advised users to save important content via copying or screenshots before the deadline. This strategic retreat from C-end agent features signals a broader market shift. Compliance capability and sustainable business models are replacing user scale and feature richness as the new core competitive dimensions. Tongyi Qianwen's recent move to fully open its Agent and Skill platforms to third-party enterprises and developers further underscores a strategic pivot from low-value C-end services to high-value B-end enterprise scenarios.

marsbit41m ago

Doubao and Qwen Will Discontinue Agent Functionality on July 15

marsbit41m ago

Why Did Codex and ChatGPT Merge? What's Next for Codex? OpenAI Core Leader Answers Everything

In 2026, OpenAI's Codex saw explosive growth, with weekly active users surging over 5x to 5 million since January, driven largely by the February launch of its desktop app. Codex desktop lead Andrew Ambrosino explains key shifts behind its evolution. A core change is the inversion of development costs: implementation is now cheap, while curation and taste—judging which of many AI-generated prototypes is valuable—have become the new scarcities. Ambrosino defines taste as a blend of aesthetics, systems thinking, direction, and semantic coherence in interaction. He notes AI still struggles with design because evaluating it requires human cultural context and abstract reasoning about how components relate—capabilities beyond current models. Timing is critical: the same Codex app would have failed months earlier; success hinges on the model's capabilities at launch. Roles are blurring within his team, with engineers, designers, and PMs overlapping significantly. However, Ambrosino cautions against eliminating specialized roles entirely, as each field retains deep expertise. On AI-assisted development, the focus has shifted from measuring code written by AI to distinguishing between supervised and unsupervised generation. A current challenge is teaching models to simplify code, not just add complexity. The merger of Codex and ChatGPT stems from observed user behavior: non-developers adopted Codex for general knowledge work despite its developer-centric interface. This revealed a collapsing boundary between specialized tools and universal assistants. The vision is a "home base" that orchestrates tasks across external professional tools (like Excel or Premiere Pro) via connectors, rather than rebuilding everything internally. An internal example showed Codex helping edit video by interacting with Premiere Pro's files and even writing a plugin for it. The future direction is a unified, extensible platform that serves as a central hub for automating and managing work across any specialized tool the user employs.

marsbit52m ago

Why Did Codex and ChatGPT Merge? What's Next for Codex? OpenAI Core Leader Answers Everything

marsbit52m ago

Trading

Spot

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of AI (AI) are presented below.

活动图片