Original source: OKX

AI tracks are undergoing an evolution from speculation to practical application.

Early AI Meme tokens took advantage of AI hotspots to explode, but now more functional AI trading tools, intelligent investment research, and on-chain AI executors are emerging. From AI-driven on-chain sniper strategies, to AI Agents ‘autonomous execution of on-chain tasks, and AI-generated DeFi revenue optimization solutions, the influence of AI tracks is rapidly expanding.

But most people can see the exponential growth in the market value of AI tokens, but cannot find a coordinate system to decode their value logic. Which AI tracks have long-term vitality? Is DeFAI the best application of AI? What are the dimensions of AI project evaluation?… The latest research report of OKX Ventures deeply disintegrates the development map of AI tracks, from concept analysis, evolution process, application tracks, and project cases, hoping to bring some inspiration and thinking to everyone’s understanding of the value of AI.

This report is relatively rich in content. In order to facilitate everyone’s reading, we have divided it into two parts (top) and (bottom). This article is the “first chapter”.

1. About AI Agents

An AI Agent is an intelligent entity with the ability to perceive the environment, make decisions, and perform corresponding actions. Unlike traditional artificial intelligence systems, AI agents can think independently and invoke tools to gradually achieve specific goals, which gives them greater autonomy and flexibility when handling complex tasks.

In short, an AI agent is an agent driven by artificial intelligence technology, and its workflow includes: perception modules (collecting input), large language models (understanding, reasoning, and planning), tool calls (performing tasks), and feedback and optimization (verification and adjustment).

OpenAI defines an AI agent as a system with a large language model as the core and the ability to independently understand, perceive, plan, remember and use tools, and can automate complex tasks. Unlike traditional artificial intelligence, AI agents can gradually complete set goals through independent thinking and tool invocation.

The definition of AI Agent can be summarized into the following key elements: perception (Perception): AI Agents sense the surrounding environment through sensors, cameras or other input devices and obtain necessary information; understand and reason Reasoning, which is capable of analyzing perceived information and making complex reasoning in order to make reasonable decisions; decision-making Decision-making: Based on the analysis results, the AI Agent can formulate an action plan and select the best execution path; Action: Finally, the AI Agent will execute the formulated plan and interact with other systems by calling external tools or interfaces., achieve predetermined goals.

The working principle and process of an AI Agent usually includes the following steps: First, information input, receiving information from the environment, such as user instructions, sensor data, etc.; then, data processing, using built-in algorithms and models to process the input data, combined with its memory system (short-term and long-term memory) to understand the current state; then, plan formulation, based on the processing results, the AI Agent divides large tasks into manageable small tasks, and formulates a specific execution plan. During the execution phase, the AI Agent implements its plan and monitors the execution process by calling external APIs or tools to ensure that the task is completed as expected; finally, feedback and learning. After the task is completed, the AI Agent will conduct self-reflection and learning based on the results, thereby improving the quality of future decisions.

2. Evolution process

The evolution path of AI tokens shows the transition from the initial “MEME” phenomenon to deep technological integration. At first, many tokens relied on brief conceptual hype and social media craze to attract users ‘attention, like Internet blockbusters. However, as the market continues to mature, AI tokens are gradually developing towards more practical and high-level functions, gradually getting rid of the pure hype model, and transforming into a true blockchain financial tool and data analysis platform. We will discuss in depth how these tokens have evolved from conceptual existence to technological products with practical application value.

Stage 1: AI Meme (confusion period)

Most of the early AI tokens existed in the form of “MEME”. Tokens such as $GOAT,$ACT, and $FARTCOIN did not have practical applications or functions. Their value mainly relied on conceptual hype and market sentiment. At this stage, the purpose of the token is not yet clear, and the market and users know little about its potential. The popularity of the token relies more on social media dissemination and short-term hype, presenting a mysterious and unpredictable characteristic.

Phase 2: Socialization (Exploration Period)

As the market gradually pays attention to AI tokens, these tokens have begun to make efforts in the social field. For example, tokens such as $LUNA and $BULLY attract user participation through enhanced social features. At this stage, tokens not only exist as hype tools, but also begin to integrate into community-driven and social interactions to promote market growth. Tokens have gradually expanded from a simple “escort chat” function and began to explore functions that are closely integrated with users ‘social needs, forming more diverse social attributes.

Stage 3: Vertical field (functional deepening period)

AI tokens began to move away from simple social and hype models and explore deep application scenarios in vertical fields. Tokens such as $AIXBT and $ZEREBRO are gradually empowered by combining them with blockchain, DeFi or creative tools, making them no longer just speculative tools, but digital assets with clear functions and purposes. This stage marks the development of AI tokens in a more efficient and professional direction, gradually forming their unique market position.

Phase 3.5: Infrastructure (technology improvement period)

While token application is gradually deepening, AI tokens are beginning to focus on building a more solid technical infrastructure. The addition of tokens such as $AI16Z and $EMP further promotes the functional optimization of tokens. Tokens not only focus on economic incentives and practical functions, but also begin to attach importance to the construction of infrastructure such as cross-chain technology, decentralized applications, and hardware integration, gradually laying a technical foundation for its future sustainable development.

Stage 4: Data analysis (maturity stage)

Entering a mature stage, AI tokens have gradually stabilized in the market and have begun to incorporate more complex cryptographic investment research and analysis functions to promote the improvement of the token ecosystem and governance structure. Tokens such as TRISIG and $COOKIE are no longer simple tools; they have become part of the economic system and are widely used in high-level areas such as data analysis, community governance, and investment decisions. At this time, the functions of AI tokens have gradually improved, and they have been able to provide in-depth analysis and decision support to the market, becoming an important asset in the encryption market.

Stage 4.5: Financial application (ecological integration period)

With the further development of the DeFi field, the integration of AI tokens in financial applications has become deeper and deeper, giving birth to the emerging concept of “DeFAI”. Through artificial intelligence, DeFi’s complex operations have become easier, and ordinary users can easily participate in online financial activities. Representative tokens such as $GRIFFAIN,$ORBIT,$AIXBT, etc. have gradually formed a complete chain from basic functions to complex financial services in the market, optimizing on-chain interactions, lowering participation thresholds, and bringing more opportunities and convenience to users.

3. AI Agent Framework

1. Comparison of Web3 and Web2 data

While Web2 ‘s AI Agents are embedded in recommendation algorithms, Web3 ‘s testing ground is also breeding more AI Agent innovations. But the data shows that Web3 and Web2 projects show significant differences in contributor distribution, code submissions, and GitHub Stars. By comparing data from Web3 and Web2 projects, we can better understand the current situation of the two in terms of technological innovation, community activity, and market acceptance. Especially on the GitHub platform, the activity and popularity of these projects provide us with important indicators that help us gain insight into future technology trends and community ecological changes.

In terms of developer participation, the number of contributors to the Web2 project is significantly higher than the Web3 project. Specifically, the Web3 project has 575 contributors, while the Web2 project has 9,940 contributors, reflecting the maturity of the Web2 ecosystem and the broader developer base. The top three contributors are: Starkchain, 3,102 contributors;Informers-agents, 3,009 contributors; and Llamaindex, 1,391 contributors.

In terms of code submission distribution. The submission volume of Web2 projects is also significantly higher than that of Web3 projects. The total number of submissions for Web3 projects was 9,238, while the number for Web2 projects was as high as 40,151, indicating that Web2 projects are more active in development and have a relatively stable update frequency. The top three projects in terms of code commits are: ElipsOS led the way with 5,905 commits; closely followed by Dust with 5,602 commits; and LangChain ranked third with 5,506 commits.

Distribution of GitHub Stars. The popularity of the Web2 project on GitHub far exceeds that of the Web3 project. The Web2 project has received a total of 526,747 Stars, while the Web3 project has received 15,676 Stars. This gap reflects the widespread recognition of the Web2 project in the developer community and its long-term accumulated market influence. The top three projects in the number of Stars are: JS Agents is undoubtedly the most popular, receiving 137,534 Stars; followed closely by LangChain, ranking second with 98,184 Stars; and MetaGPT ranked third, receiving 46,676 Stars.

Overall, the Web2 project leads significantly in terms of the number of contributors and code submission frequency, demonstrating its mature and stable ecosystem. The huge developer base and continuous technological innovation have kept Web2 projects highly competitive in the market. In contrast, although the number of contributors to the Web3 project is small, some projects have outstanding performance in terms of code submission frequency, indicating that they have a stable core development team and can continue to promote project development. Although the Web3 ecosystem is relatively preliminary at present, its potential cannot be underestimated. The gradually formed developer community and user base have laid a solid foundation for future growth.

In terms of project popularity, the distribution of GitHub Stars reveals the important role of JavaScript and Python in the development of AI proxy frameworks. JS Agents and LangChain are the most popular projects, showing that the trend of combining AI with cryptocurrencies is receiving widespread attention. Although the number of Stars in the Web3 project is much lower than that in the Web 2 project, some Web3 projects such as MetaGPT still perform well and win the recognition of developers. Overall, although the Web3 project is in the catch-up stage, its position in the future market is expected to steadily increase as the technology further matures and ecological expansion.

(2) Mainstream blockchain AI Agent framework

Data source: www.aiagenttoolkit.xyz/#frameworks

(3) Challenges faced by existing blockchain AI Agent frameworks

The “dimension reduction attack” of large manufacturers ‘competing products. Technology giants such as OpenAI, Google, and Microsoft are rapidly launching official-level multi-tool agents. With their strong financial and technological advantages, they can occupy the market and marginalize start-up frameworks at any time. By deeply integrating large language models (LLMs), cloud services and tool ecosystems, these large manufacturers can provide comprehensive and efficient solutions, putting small and medium-sized frameworks under greater competitive pressure and greatly squeezing their living space.

Stability and maintainability are not enough. At present, all AI agents generally face high error rates and “illusion” problems, especially when calling the model in multiple rounds, which are prone to infinite loops or compatibility bugs. Once an agent is required to perform multiple subtasks, these errors are often amplified layer by layer, resulting in system instability. For enterprise applications that require a high degree of reliability, these frameworks are currently unable to provide sufficient stability and production-level guarantees, limiting their widespread application in real business environments.

Performance and costs remain high. Agentization processes usually require a large number of inference calls (such as loop self-checking, tool functions, etc.). If the underlying layer relies on large models such as GPT-4, it will face high call costs and often fail to meet the need for rapid response. Although some frameworks try to combine open source models for local reasoning to reduce costs, this method still relies on strong computing power, and the quality of reasoning results is difficult to stabilize, requiring continuous optimization by professional teams to ensure system reliability and performance.

Insufficient development ecology and flexibility. At present, these AI proxy frameworks lack unified standards in terms of development language and extensibility, causing developers to face certain confusion and limitations when choosing. For example, Eliza uses TypeScript, which is easy to get started, but is poorly scalable in high-complexity scenarios;Rig uses Rust, which has excellent performance but a high learning threshold;ZerePy (ZEREBRO) is based on Python and is suitable for creative generation class applications, but its functions are relatively limited. Other frameworks such as AIXBT and Griffain are more focused on specific blockchain or vertical domain applications, and market verification still takes time. Developers often need to make trade-offs between these frameworks between ease of use, performance and multi-platform adaptation, affecting their flexibility and development potential in wider applications.

Security and compliance risks. When multi-agent systems access external APIs, perform critical transactions, or make automated decisions, they are prone to security risks such as unauthorized calls, privacy leaks, or vulnerable operations. Many frameworks are not yet complete in handling security policies and audit records, especially in enterprise or financial application scenarios, where these issues are extremely prominent and difficult to meet strict compliance requirements. This makes the system likely to face great legal risks and data security challenges when actually deployed.

In view of the above problems, many practitioners believe that the current AI Agent framework may be further squeezed under the pressure of “the next technological breakthrough” or “big factory integration plan.” However, there are also views that start-up frameworks can still play unique value in specific areas, such as on-chain scenarios, creative generation or community plug-in docking. As long as breakthroughs can be made in reliability, cost control and ecological construction, these frameworks can still find feasible development paths outside the ecology of large factories. Overall, how to solve the two major problems of “high cost, error-prone” and “achieving multi-scenario flexibility” will be key challenges faced by all AI Agent frameworks.

3. Development direction of AI Agent

The popularity of multimodal AI

With the rapid development of technology, multimodal AI is gradually becoming a key driving force in various industries. Multimodal AI can process multiple data forms such as text, images, video and audio, making it show great potential in multiple fields. Especially in the medical field, by integrating medical records, imaging data and genomic information, multimodal AI can support the implementation of personalized medicine and help doctors more accurately tailor treatment plans to patients. In retail and manufacturing, with this technology, AI can optimize production processes, improve efficiency, and enhance customer experience, thereby enhancing the competitiveness of enterprises. With the improvement of data and computing capabilities, multimodal AI is expected to play a transformative role in more industries, promoting rapid iteration and application expansion of technology.

Specific intelligence and autonomous intelligence

Embodied AI refers to the way artificial intelligence systems understand and adapt to the environment through perception and interaction with the physical world. This technology will greatly change the development direction of robots and lay the foundation for their popularity in autonomous driving, smart cities and other application scenarios. 2025 is regarded as the “first year of embodied intelligence”, and this technology is expected to be widely used in many fields. By giving robots the ability to perceive, understand and make independent decisions, embodied intelligence will promote the deep integration of the physical world and the digital world, thereby improving productivity and promoting intelligent development in all walks of life. Whether it’s in personal assistants, autonomous vehicles, or smart factories, embodied intelligence will change the way people interact with machines.

The rise of AI agents

AI agents refer to artificial intelligence systems that are capable of completing complex tasks independently. Such AI agents are transforming from early simple query response tools to more advanced autonomous decision-making systems, and are widely used in fields such as business process optimization, customer service, and industrial automation. For example, AI agents can independently handle customers ‘consultation requests, provide personalized services, and even make optimization decisions. In industrial automation, AI agents can monitor the operating status of equipment, predict failures, and make adjustments or repairs before problems occur. As AI agents gradually mature, their application in various industries will become more in-depth and become an important tool to improve efficiency and reduce costs.

Application of AI in scientific research

The introduction of AI is accelerating the progress of scientific research, especially in the field of complex data analysis. AI4S (AI for Science) has become a new research trend. Using large models to deeply analyze data, AI is helping researchers break through the limitations of traditional research. In fields such as biomedicine, materials science and energy research, the application of AI is driving breakthroughs in basic science. A notable example is AlphaFold, which solves long-standing problems that have plagued scientists by predicting the structure of proteins and greatly promotes the progress of biomedical research. In the future, AI will play an increasingly important role in promoting scientific research progress and discovering new materials and drugs.

AI safety and ethics

With the popularization of AI technology, AI safety and ethical issues are gradually becoming the focus of global attention. The transparency, fairness and potential security risks of AI systems in decision-making have triggered a lot of discussion. In order to ensure the sustainable development of AI technology, companies and officials are stepping up efforts to establish a sound governance framework to effectively manage its risks while promoting technological innovation. Especially in areas such as automated decision-making, data privacy and autonomous systems, how to balance technological progress with social responsibility has become the key to ensuring the positive impact of AI technology. This is not only a challenge for technological development, but also an important issue at the moral and legal levels, affecting the role and status of AI in future society.

In the “second part” of the report, we will introduce in detail the application and table items of AI Agents, and provide an evaluation framework. Stay tuned.

disclaimer

This content is for reference only and does not constitute and should not be regarded as (i) investment advice or recommendation,(ii) an offer or solicitation to buy, sell or hold digital assets, or (iii) financial, accounting, legal or tax advice. We do not guarantee the accuracy, completeness or usefulness of this information. Digital assets (including stablecoins and NFTs) are subject to market fluctuations, involve high risks, and may devalue or even become worthless. You should carefully consider whether trading or holding digital assets is suitable for you based on your financial situation and risk tolerance. Please consult your legal/tax/investment professional for your specific situation. Not all products are available in all regions. For more details, please refer to OKX Terms of Service and Risk Disclosure & Disclaimer. OKX Web3 Mobile Wallet and its derivative services are subject to separate terms of service. It is your own responsibility to understand and comply with applicable local laws and regulations.

This article is from a submission and does not represent the views of BlockBeats