In-depth analysis of Multi-Agent: Will Web3 and AI eventually complement each other?

If AIGC has opened the intelligent era of content generation, then AI Agent has the opportunity to truly productize the capabilities of AIGC.

AI Agent is like a more concrete all-round employee, known as the primary form of artificial intelligence robots, which can observe the surrounding environment, make decisions, and take actions automatically like humans.

Bill Gates once said bluntly, "Mastering AI Agent is the real achievement. By then, you will no longer need to search for information on the Internet yourself." Authoritative experts in the field of AI also have high hopes for the prospects of AI Agent. Microsoft CEO Satya Nadella once predicted that AI Agent will become the main way of human-computer interaction, able to understand user needs and actively provide services. Professor Andrew Ng also predicted that in the future work environment, humans and AI Agents will collaborate in a closer way to form an efficient working model and improve efficiency.

AI Agent is not only a product of technology, but also the core of future life and work methods.

This reminds us that when Web3 and blockchain first caused widespread discussion, people often used the word "disruption" to describe the potential of this technology. Looking back over the past few years, Web3 has gradually developed from the initial ERC-20 and zero-knowledge proof to DeFi, DePIN, GameFi, etc. that are integrated with other fields.

If Web3 and AI, two popular digital technologies, are combined, will there be a 1+1>2 effect? ​​Can Web3 AI projects, which are increasingly raising funds, bring new use case paradigms to the industry and create new real needs?

AI Agent: The most ideal intelligent assistant for humans

Where is the imagination of AI Agent? A high-scoring answer is widely circulated on the Internet, "A large language model can only compile a greedy snake, while an AI Agent can compile an entire King of Glory." It sounds exaggerated, but it is not an exaggeration.

Agent, usually translated as "intelligent body" in China. This concept was proposed by Minsky, the "father of artificial intelligence" in his book "The Society of Thinking" published in 1986. Minsky believed that some individuals in society can reach a solution to a problem after negotiation, and these individuals are agents. For many years, agents have been the cornerstone of human-computer interaction. From Microsoft's editing assistant Clippy to Google Docs' automatic suggestions, these early forms of agents have shown the potential for personalized interaction, but their ability to handle more complex tasks is still limited. It was not until the emergence of large language models (LLMs) that the true potential of agents was tapped.

In May this year, Professor Andrew Ng, an authoritative scholar in the field of AI, shared a speech on AI Agents at the Sequoia AI event in the United States. In it, he showed a series of experiments done by his team:

Let AI write some code and run it, and compare the results obtained by different LLMs and workflows. The results are as follows:

• GPT-3.5 model: 48% accuracy

• GPT-4 model: 67% accuracy

• GPT-3.5 + Agent: higher than the performance of the GPT-4 model

• GPT-4 + Agent: much higher than the GPT-4 model, very outstanding

In contrast, the AI ​​Agent workflow is as follows:

First, let the LLM write an article outline. If necessary, search the content on the Internet for research and analysis, output a draft, and then read the draft and think about how to optimize it. This cycle repeats and iterates many times, and finally outputs a high-quality article with rigorous logic and the lowest error rate.

We can find that the difference between AI Agent and LLM is that the interaction between LLM and humans is based on prompts. AI Agent only needs to set a goal, and it can think independently and act on the goal. According to the given task, each step of the plan is broken down in detail, relying on feedback from the outside world and independent thinking, and creating prompts for itself to achieve the goal.

Therefore, OpenAI defines AI Agent as: A system that uses LLM as its brain, has the ability to autonomously understand, perceive, plan, remember, and use tools, and can automatically perform complex tasks.

When AI changes from a tool to a subject that can use tools, it becomes an AI Agent. This is exactly why AI Agent can become the most ideal intelligent assistant for humans. For example, based on the user's historical online interactions, AI Agent can understand and remember the user's interests, preferences, and daily habits, identify the user's intentions, proactively make suggestions, and coordinate multiple applications to complete tasks.

Just as Gates envisioned, in the future we no longer need to switch to different applications for different tasks. We only need to use ordinary language to tell computers and mobile phones what we want to do. According to the data that users are willing to share, AI Agent will provide personalized responses.

Single-person unicorn companies are becoming a reality

AI Agent can also help companies create a new intelligent operation model with "human-machine collaboration" as the core. More and more business activities will be completed by AI, and humans only need to focus on corporate vision, strategy and key path decision-making.

As OpenAI CEO Sam Altman once mentioned in an interview, with the development of AI, we are about to enter the era of "single-person unicorns", that is, companies founded by a single person and valued at $1 billion.

It sounds like fantasy, but with the help of AI Agents, this view is becoming a reality.

Let's assume that we are now starting a technology startup. According to the traditional method, I obviously need to hire software engineers, product managers, designers, marketing personnel, sales and finance personnel, each of whom has their own responsibilities but I coordinate them all.

If I use AI Agents, I may not even need to hire employees.

• Devin — Automated Programming

Instead of software engineers, I might use Devin, an AI software engineer that has exploded this year, which can help me complete all front-end and back-end work.

Developed by Cognition Labs, Devin is known as "the world's first AI software engineer." It can independently complete the entire software development work, analyze problems, make decisions, write code, and fix errors independently, all of which can be executed autonomously. It greatly reduces the workload of developers. Devin received $196 million in financing in just half a year, and its valuation quickly soared to billions of dollars. Investors include well-known venture capital companies such as Founders Fund and Khosla Ventures.

Although Devin has not yet launched a public version, we can get a glimpse of its potential from Cursor, another recently popular Web2 product. It can do almost all the work for you, turning a simple idea into functional code in a few minutes. You only need to give orders and you can "sit back and enjoy the results". It is reported that an eight-year-old child, without any programming experience, actually used Cursor to complete the code work and built a website.

• Hebbia — File Processing

Instead of a product manager or financial staff, I might choose Hebbia, which can help me organize and analyze all documents.

Unlike Glean, which focuses on document search within the enterprise, Hebbia Matrix is ​​an enterprise-level AI Agent platform that uses multiple AI models to help users efficiently extract, structure, and analyze data and documents, thereby promoting the improvement of enterprise productivity. It is impressive that Matrix can process millions of documents at a time.

Hebbia completed a $130 million Series B round in July this year, led by a16z, with participation from well-known investors such as Google Ventures and Peter Thiel.

• Jasper AI — Content Generation

Instead of social media operations and designers, I might choose Jasper AI, which can help me generate content.

Jasper AI is an AI Agent writing assistant designed to help creators, marketers and companies simplify the content generation process and improve productivity and creative efficiency. Jasper AI can generate a variety of content types according to the style required by the user, including blog posts, social media posts, advertising copy and product descriptions. And generate pictures based on the user's description to provide visual assistance for text content.

Jasper AI has received $125 million in financing and reached a valuation of $1.5 billion in 2022. According to statistics, Jasper AI has helped users generate more than 500 million words, becoming one of the most widely used AI writing tools.

• MultiOn — Web Automation

Instead of an assistant, I might choose MultiOn to help me manage daily tasks, schedule, set reminders, and even plan business trips, automatically book hotels, and automatically arrange online rides.

MultiOn is an automated web task AI agent that can help perform tasks autonomously in any digital environment, such as helping users complete personal tasks such as online shopping and appointments, improve personal efficiency, or help users simplify daily affairs and improve work efficiency.

• Perplexity — Search, Research

To replace the researcher, I might choose Perplexity, which is used by Nvidia's CEO every day.

Perplexity is an AI search engine that can understand the user's questions, break down the questions, and then search and integrate the content to generate reports to provide users with clear answers.

Perplexity is suitable for various user groups. For example, students and researchers can simplify the information retrieval process when writing and improve efficiency; marketers can obtain reliable data to support marketing strategies.

The above content is only imagination. The real ability and level of these AI Agents are not enough to replace elite talents in all walks of life. As Li Bojie, co-founder of Logenic AI, said, the current LLM capabilities are only at the entry-level level, far from the expert level. At this stage, AI Agent is more like an employee who works faster but is not very reliable.

However, these AI Agents, with their respective strengths, are helping existing users improve efficiency and convenience in a variety of scenarios.

Not only limited to technology companies, all walks of life can benefit from the wave of AI Agents. In the field of education, AI Agents can provide personalized learning resources and tutoring based on students' learning progress, interests and abilities; in the field of finance, AI Agents can help users manage personal finances, provide investment advice, and even predict stock trends; in the medical field, AI Agents can help doctors diagnose diseases and formulate treatment plans; in the field of e-commerce, AI Agents can also serve as intelligent customer service, automatically answering user inquiries, handling order issues and return requests through natural language processing and machine learning technology, thereby improving customer service efficiency.

Multi-Agent: The Next Step for AI Agent

In the previous section about the idea of ​​a single-person unicorn company, a single AI Agent faces limitations when handling complex tasks and is difficult to meet actual needs. When using multiple AI Agents, since these AI Agents are based on heterogeneous LLMs, collective decision-making is difficult and their capabilities are limited, humans are still needed to act as dispatchers between these independent AI Agents to coordinate these AI Agents serving different application scenarios to work. This gave rise to the rise of "Multi-Agent (multi-agent framework)".

Complex problems often require the integration of multiple aspects of knowledge and skills, but a single AI Agent has limited capabilities and is unable to cope with them. By organically combining AI Agents with different capabilities, the Multi-Agent system allows AI Agents to play their respective strengths, learn from each other's strengths and make up for each other's weaknesses, thereby solving complex problems more effectively.

This is very similar to our actual workflow or organizational structure: a leader assigns tasks, people with different abilities are responsible for different tasks, the results of each process are given to the next process, and finally the final task results are obtained.

In the implementation process, lower-level AI agents perform their respective tasks, while higher-level AI agents assign tasks and supervise their completion.

Multi-Agent can also simulate our human decision-making process. Just like we will consult with others when we encounter problems, multiple AI agents can also simulate the behavior of collective decision-making and provide us with better information support. For example, AutoGen developed by Microsoft meets this requirement:

• It can create AI agents with different roles. These AI agents have basic conversation capabilities and can generate replies based on the received messages.

• Use GroupChat to create a group chat environment with multiple AI Agents. In this GroupChat, there is an AI Agent with the role of administrator to manage the chat records, speaker order, and termination of speech of other AI Agents.

If applied to the concept of a single-person unicorn company, we can create several AI Agents with different roles through the Multi-Agent architecture, such as project managers, programmers, or supervisors. Tell them our goals and let them think of ways. We just need to listen to their reports. If we have opinions or find something they did wrong, we will let them change it until they are satisfied.

Compared with a single AI Agent, Multi-Agent can achieve:

• Scalability:Handle larger problems by increasing the number of AI Agents, with each AI Agent handling a part of the task, allowing the system to scale as demand grows.

• Parallelism:Naturally supports parallel processing, and multiple AI Agents can work on different parts of the problem at the same time, accelerating problem solving.

• Decision Improvement:Enhance decision making by aggregating the insights of multiple AI Agents, as each AI Agent has its own perspective and expertise.

With the continuous advancement of AI technology, it is conceivable that the Multi-Agent framework will play a greater role in more industries and promote the development of various new AI-driven solutions.

AI Agent wind blows to Web3 

Stepping out of the laboratory, AI Agent and Multi-Agent have a long way to go.

Leaving aside Multi-Agent, even the most advanced single AI Agent at present still has a clear upper limit on the computing resources and computing power required at the physical level, and cannot be infinitely expanded. Once faced with extremely complex and computationally intensive tasks, AI Agent will undoubtedly encounter a computing bottleneck and its performance will be greatly reduced.

Furthermore, AI Agent and Multi-Agent system are essentially a centralized architecture mode, which determines that it has a very high risk of single failure. More importantly, the monopoly business model based on closed-source large models by companies such as OpenAI, Microsoft, and Google seriously threatens the survival environment of independent and single AI Agent startups, making it impossible for AI Agents to smoothly use huge corporate private data to make them smarter and more efficient. AI Agents are in urgent need of a democratic collaborative environment so that truly valuable AI Agents can serve a wider range of people in need and create greater value for society.

Finally, although AI Agent is closer to the industry than LLM, its development is based on LLM, and the current large model track is characterized by high technical barriers, high capital investment, and immature business models. AI Agents usually find it difficult to obtain financing for continuous updates and iterations.

The Multi-Agent paradigm is an excellent angle for Web3 to help AI. Many Web3 development teams are already investing in research and development to provide solutions in these areas.

AI Agent and Multi-Agent systems usually require a lot of computing resources to make complex decisions and process tasks. Web3 can build a decentralized computing power market through blockchain and decentralized technology, so that computing power resources can be distributed and utilized more fairly and efficiently on a global scale. Web3 projects such as Akash, Nosana, Aethir, IO.net, etc. can provide computing power for AI Agent decision-making and reasoning.

Traditional AI systems are often centrally managed, which causes AI Agents to face single point failures and data privacy issues. The decentralized nature of Web3 can make Multi-Agent systems more decentralized and autonomous. Each AI Agent can run independently on different nodes and autonomously execute the needs of users, which enhances robustness and security. Establishing incentive and punishment mechanisms for pledgers and delegators through mechanisms such as PoS and DPoS can promote the democratization of single AI Agent or Multi-Agent systems.

In this regard, GaiaNet, Theoriq, PIN AI, and HajimeAI have all made very cutting-edge attempts.

• Theoriq is a project serving "AI for Web3". It hopes to establish the calling and economic system of AI Agents through Agentic Protocol, popularize the development of Web3 and many functional scenarios, and provide verifiable model reasoning capabilities for Web3 dApp.

• GaiaNet is a node-based AI Agent creation and deployment environment. It takes the protection of the intellectual property rights and data privacy of experts and users as its starting point to counter the centralized OpenAI GPT Store.

• HajimeAI, on the basis of both, focuses on the establishment of AI Agent workflows in actual needs and the intelligence and automation of the intent itself, echoing the "personalization of AI intelligence" mentioned by PIN AI.

• At the same time, Modulus Labs and ORA Protocol have made progress in the algorithm direction of zkML and opML of AI Agent respectively.

Finally, the development and iteration of AI Agent and Multi-Agent systems often require a lot of financial support, and Web3 can help potential AI Agent projects obtain valuable early support through the characteristics of front-end liquidity.

Both Spectral and HajimeAI have proposed product ideas to support the issuance of AI Agent assets on the chain: by issuing tokens through IAO (Initial Agent Offering), AI Agent can directly obtain funds from investors and become a member of DAO governance, providing investors with the opportunity to participate in project development and share future benefits. Among them, HajimeAI's Benchmark DAO hopes to organically combine decentralized AI Agent scoring and AI Agent asset issuance through crowdfunding and token incentives, and create a closed loop of AI Agent financing and cold start based on Web3, which is also a relatively novel attempt.

The AI ​​Pandora's box has been opened, and everyone in it is both excited and confused. No one knows whether it is an opportunity or a reef under the craze. Today, all walks of life are no longer in the era of PPT financing. No matter how cutting-edge the technology is, it can only realize its value through implementation. The future of AI Agent is destined to be a long marathon, and Web3 is ensuring that it will not be out of the game.