Written by: 0XNATALIE
Since the second half of this year, the topic of AI Agents has been rising steadily. Initially, the AI chatbot terminal of truths gained widespread attention for its humorous posts and replies on X (similar to 'Robert' on Weibo) and received a $50,000 grant from a16z founder Marc Andreessen. Inspired by its content, someone created the GOAT token, which saw a rise of over 10,000% within just 24 hours. The topic of AI Agents subsequently attracted the attention of the Web3 community. Later, the first decentralized AI trading fund based on Solana, ai16z, was launched, introducing the AI Agent development framework Eliza and sparking a token competition. However, the community's understanding of the AI Agent concept remains unclear: what is the core of AI Agents? How are they different from Telegram trading bots?
Working principle: Perception, reasoning, and autonomous decision-making
The AI Agent is an intelligent agent system based on large language models (LLM), capable of perceiving the environment, making reasoning decisions, and completing complex tasks by calling tools or executing operations. Workflow: Perception module (acquiring input) → LLM (understanding, reasoning, and planning) → Tool invocation (executing tasks) → Feedback and optimization (verification and adjustment).
Specifically, the AI Agent first acquires data from the external environment through the perception module (such as text, audio, images, etc.) and converts it into structured information that can be processed. The LLM, as the core component, provides powerful natural language understanding and generation capabilities, acting as the system's 'brain'. Based on the input data and existing knowledge, the LLM performs logical reasoning, generating possible solutions or formulating action plans. Subsequently, the AI Agent completes specific tasks by invoking external tools, plugins, or APIs and validates and adjusts the results based on feedback, forming a closed-loop optimization.
In the application scenarios of Web3, what is the difference between AI Agents and Telegram trading bots or automated scripts? Taking arbitrage as an example, users wish to perform arbitrage trading when profits exceed 1%. In Telegram trading bots that support arbitrage, users set up trading strategies with profits greater than 1%, and the bot begins execution. However, when the market fluctuates frequently, and arbitrage opportunities change continuously, these bots lack risk assessment capabilities, executing arbitrage as long as the profit condition is met. In contrast, AI Agents can automatically adjust their strategies. For example, when a trade's profit exceeds 1%, but data analysis assesses the risk to be too high, potentially causing losses due to sudden market changes, it will decide not to execute that arbitrage.
Therefore, the AI Agent possesses self-adaptability, with its core advantage being the ability to self-learn and make autonomous decisions. Through interaction with the environment (such as the market, user behavior, etc.), it adjusts its behavioral strategies based on feedback signals, continuously improving task execution effectiveness. It can also make real-time decisions based on external data and continuously optimize decision-making strategies through reinforcement learning.
Does this sound a bit like a solver under the intent framework? The AI Agent itself is also a product based on intent, with the biggest difference from solvers under the intent framework being that solvers rely on precise algorithms with mathematical rigor, while AI Agent's decisions depend on data training, often requiring continuous trial and error during the training process to approach the optimal solution.
Mainstream frameworks for AI Agents
The AI Agent framework is the infrastructure for creating and managing intelligent agents. Currently, in Web3, popular frameworks include ai16z's Eliza, zerebro's ZerePy, and Virtuals' GAME.
Eliza is a multifunctional AI Agent framework built using TypeScript, supporting operation across multiple platforms (such as Discord, Twitter, Telegram, etc.). Through complex memory management, it can remember previous conversations and contexts, maintaining stable and consistent personality traits and knowledge responses. Eliza adopts a RAG (Retrieval Augmented Generation) system, enabling access to external databases or resources for generating more accurate answers. Additionally, Eliza integrates TEE plugins, allowing deployment in a TEE environment to ensure data security and privacy.
GAME is a framework that empowers and drives AI Agents to make autonomous decisions and actions. Developers can customize the behavior of the agents according to their needs, extend their functionalities, and provide tailored operations (such as social media posting, replying, etc.). Different functionalities in the framework, such as the agent's environmental position and tasks, are divided into multiple modules for easy configuration and management by developers. The GAME framework divides the decision-making process of AI Agents into two levels: High-Level Planning (HLP) and Low-Level Planning (LLP), each responsible for different levels of tasks and decisions. High-Level Planning sets the overall goals and task planning for agents, making decisions based on objectives, personality, background information, and environmental states, determining task priorities. Low-Level Planning focuses on execution, translating high-level planning decisions into specific operational steps, selecting appropriate functions and methods.
ZerePy is an open-source Python framework for deploying AI Agents on X. This framework integrates LLMs provided by OpenAI and Anthropic, allowing developers to build and manage social media agents that automate actions such as posting tweets, replying to tweets, liking, etc. Each task can be assigned different weights based on its importance. ZerePy provides a simple command-line interface (CLI) for developers to quickly start and manage agents. At the same time, the framework also offers Replit (an online code editing and execution platform) templates, enabling developers to quickly get started with ZerePy without complex local environment configurations.
Why does the AI Agent face FUD?
The AI Agent appears intelligent and can lower entry barriers and enhance user experience, but why is there FUD in the community? The reason is that the AI Agent is essentially still just a tool; it cannot complete the entire workflow at present and can only improve efficiency and save time at certain nodes. Moreover, at this current development stage, the role of AI Agents mainly focuses on helping users issue MeMe with one click and manage social media accounts. The community jokingly says, 'assets belong to Dev, liabilities belong to AI.'
However, just this week, aiPool launched the AI Agent for token presales, utilizing TEE technology for trustlessness. The wallet private keys of this AI Agent are dynamically generated in the TEE environment, ensuring security. Users can send funds (e.g., SOL) to the wallet controlled by the AI Agent, which then creates tokens based on set rules and launches a liquidity pool on a DEX, while distributing tokens to eligible investors. The entire process does not rely on any third-party intermediaries and is fully autonomously completed by the AI Agent in the TEE environment, avoiding common rug pull risks in DeFi. It is evident that AI Agents are gradually evolving. I believe that AI Agents can help users lower barriers and enhance experiences, even if it is just simplifying part of the asset issuance process; it is meaningful. However, from a macro Web3 perspective, AI Agents, as off-chain products, currently only serve as auxiliary tools for smart contracts, so there is no need to overhype their capabilities. Due to the lack of significant wealth effect narratives besides MeMe in the second half of this year, the hype around AI Agents, centered on MeMe, is also normal. Relying solely on MeMe cannot sustain long-term value, so if AI Agents can bring more innovative gameplay to trading processes and provide tangible landing value, they may develop into a universal infra tool.
