Some friends say that the continuous decline of web3 AI Agent targets like #ai16z and $arc is caused by the recently popular MCP protocol? At first glance, one might be confused, WTF does it have anything to do with that? But upon further reflection, there is indeed a certain logic: the valuation logic of existing web3 AI Agents has changed, and the narrative direction and product landing routes need to be adjusted urgently. Below, I will share my personal views:

1) MCP (Model Context Protocol) is an open standardized protocol aimed at seamlessly connecting various AI LLM/Agents to different data sources and tools, equivalent to a plug-and-play USB 'universal' interface, replacing the previously required end-to-end 'specific' packaging methods.

In simple terms, there were obvious data silos between AI applications, and to achieve interoperability between Agents/LLMs, each one needed to develop corresponding API interfaces, which not only complicated the operational process but also lacked bidirectional interaction functionality, usually having relatively limited model access and permission restrictions.

The emergence of MCP equals providing a unified framework, allowing AI applications to escape the past state of data silos and realize the possibility of 'dynamic' access to external data and tools, significantly reducing development complexity and integration efficiency, as well as in areas such as automated task execution, real-time data querying, and cross-platform collaboration. Speaking of this, many people immediately think, if using the Manus integration of multi-Agent collaborative innovation can promote the multi-Agent collaboration of the MCP open-source framework, wouldn't that be invincible?

That's right, Manus + MCP is the key to this round of impact on web3 AI Agents.

2) However, it is astonishing that both Manus and MCP are frameworks and protocol standards aimed at web2 LLM/Agents, addressing the issues of data interaction and collaboration between centralized servers, whose permissions and access control still rely on the 'active' opening of each server node; in other words, it is merely an open-source tool attribute.

Logically, it completely contradicts the central ideas pursued by web3 AI Agents, such as 'distributed servers, distributed collaboration, distributed incentives', etc. How can a centralized Italian cannon blow up a decentralized fortress?

The reason lies in the fact that the first phase of web3 AI Agents has become too 'web2-like'. On one hand, this comes from many teams having web2 backgrounds and lacking a thorough understanding of the native needs of web3. For example, the ElizaOS framework was initially a packaging framework that helped developers quickly deploy AI Agent applications, precisely integrating platforms like Twitter, Discord, and some API interfaces from OpenAI, Claude, DeepSeek, etc., while appropriately packaging some Memory, Character general frameworks to help developers rapidly develop and finalize AI Agent applications. But to be precise, what is the difference between this service framework and the open-source tools of web2? What differentiated advantages does it have?

Uh, is the advantage just having a set of Tokenomics incentive mechanisms? Then using a framework that can be completely replaced by a web2 framework, incentivizing a batch of AI Agents that exist mainly for launching new tokens? Terrifying... Following this logic, you can roughly understand why Manus + MCP can impact web3 AI Agents. Because many web3 AI Agent frameworks and services only address the rapid development and application needs similar to web2 AI Agents, but they cannot keep up with the innovation speed of web2 in terms of technical services, standards, and differentiated advantages, the market/capital has reassessed and repriced the previous batch of web3 AI Agents.

3) Speaking of this, the general problem should have found its crux, but how should we break through? There is only one way: focus on creating web3 native solutions, because the operation and incentive architecture of distributed systems are the absolute differentiated advantages belonging to web3?

Taking distributed cloud computing, data, algorithms, and other service platforms as an example, on the surface, it seems that this type of computing power and data aggregated under the guise of idle resources cannot meet the needs for engineering innovation in the short term. However, at a time when a large number of AI LLMs are competing in centralized computing for performance breakthroughs, a service model that touts 'idle resources, low cost' will naturally be looked down upon by web2 developers and VC teams.

But once web2 AI Agents have passed the stage of competing on performance innovation, they will inevitably pursue vertical application scenario expansion and fine-tuning model optimization directions, at which point the advantages of web3 AI resource services will truly manifest. In fact, when web2 AI, which has climbed to a certain stage via resource monopolization, finds it difficult to revert to a mindset of surrounding cities with rural areas, breaking down various segmented scenarios one by one, that will be the time when the excess web2 AI developers and web3 AI resources band together to exert force.

Therefore, the opportunity space for web3 AI Agents has become quite clear: before the web3 AI resource platform has overflowed with web2 developer demand clients, explore and implement a feasible solution and path that is indispensable for a non-web3 distributed architecture. In fact, apart from the quick deployment + multi-Agent collaboration communication framework of web2 + Tokenomic narrative, there are many web3 native innovative directions worth exploring.

For example, equipped with a distributed consensus collaboration framework, considering the characteristics of off-chain computation of large LLM models + on-chain state storage, numerous adaptable components are needed.

1) A decentralized DID identity verification system that allows Agents to have verifiable on-chain identities, similar to the unique addresses generated by virtual machines for smart contracts, primarily for subsequent state tracking and recording.

2) A decentralized Oracle oracle system, primarily responsible for the trustworthy acquisition and verification of off-chain data. Unlike previous Oracles, this oracle, tailored for AI Agents, may also need to form a composite architecture that includes data collection layers, decision consensus layers, execution feedback layers, and multiple Agents to ensure that the data required by Agents on-chain and the off-chain computation and decision-making can be accessed in real-time.

3) A decentralized storage DA system, due to the uncertainty of the knowledge base state when AI Agents operate and the temporary nature of the reasoning process, requires a system that records and stores the key state libraries and reasoning paths behind LLMs in a distributed storage system, providing a cost-controllable data proof mechanism to ensure the availability of data during public chain verification.

4) A zero-knowledge proof (ZKP) privacy computing layer can link to privacy computing solutions including TEE, FHE, etc., to achieve real-time privacy computing + data proof verification, allowing Agents to have a wider range of vertical data sources (medical, financial), thus enabling more specialized customized service Agents to emerge on top.

5) A cross-chain interoperability protocol, somewhat similar to the framework defined by the MCP open-source protocol, with the difference being that this interoperability solution requires a relay and communication scheduling mechanism that adapts to Agent operation, transmission, and verification, capable of completing asset transfers and state synchronization issues for Agents across different chains, especially involving complex states such as Agent context, Prompt, knowledge base, Memory, etc.

…

In my opinion, the focus of truly conquering web3 AI Agents should be on how to make the 'complex workflows' of AI Agents and the 'trust verification flows' of blockchains align as closely as possible. As for these incremental solutions, whether they are upgraded iterations from existing old narrative projects or newly forged projects in the AI Agent narrative track, both possibilities exist.

This is the direction that web3 AI Agents should strive to build, aligning with the basic fundamentals of the innovative ecosystem under the macro narrative of AI + Crypto. If relevant innovations and differentiated competitive barriers cannot be established, then every disturbance in the web2 AI track may cause upheaval in the web3 AI space.