On January 16, 2025, the highly anticipated Solana AI Hackathon officially announced its results. This 15-day AI hackathon attracted the enthusiastic participation of more than 400 projects. After fierce competition, the DeFAI project The Hive (BUZZ) won the championship and received a prize of US$60,000.
The Hive is a modular, interoperable DeFi Agent network built on Solana. The Hive is based on the Swarm architecture and can automatically process complex DeFi operations such as transactions, pledges, and loans through user chat commands. The Hive was developed by @jsonhedman, who also won the first place in the Microsoft Generative AI Hackathon, the first place in the Microsoft AI Classroom Hackathon, the second place in the Vanderbilt Generative AI Showcase, and the first place in the Coinbase x Polygon x Thirdweb Hack House.
The Hives popularity has once again brought the DeFAI sector to a head, with the sectors market value once reaching $2.7 billion (now it has fallen back to $800 million). DeFAI is a term coined by @danielesesta, the founder of @HeyAnonai, as well as the founder of Wonderland, Abracadabra and WAGMI. Conceptually, DeFAI is a combination of decentralized finance (DeFi) and artificial intelligence (AI), aiming to simplify the complex process of DeFi through AI technology, so that ordinary users can also easily use DeFi services.
From Intent Trading to DeFAI
The Hive is not the first DeFAI project, and the combination of DeFi and AI is not a new initiative. As early as 2023, discussions around AI and intentional transactions have become a hot topic in the blockchain community. The intention here is a set of declarative constraints that allow users to delegate transaction creation to a dedicated third-party participant network while retaining full control over the process.
An example is that Alice has 1,000 USDC on Ethereum, and she wants to use this money to buy as much BUZZ as possible on Solana. Alice first needs to connect to the Ethereum and Solana networks, which requires connecting to a node service, and can choose ZAN Node (https://zan.top/home/node-service?chInfo=ch_WZ). Then she crosses her funds from Ethereum to Solana, which requires choosing a cross-chain bridge. Fortunately, this can be achieved through CCTP, the USDC cross-chain transfer protocol officially launched by Circle. After crossing USDC to Solana, Alice now needs to decide which asset to use to buy BUZZ - because different trading pools will charge dynamic fees for different assets when buying BUZZ to balance the trading pool, which means that if the fee of USDC is too high, it may be profitable for Alice to exchange USDC for another asset (such as SOL) to buy BUZZ. At the same time, there are many SOL-BUZZ trading pools on the Solana network. For example, the commonly used DEX platforms Raydium, Orca, and Meteora all have multiple trading pools. Alice needs to compare the rates, prices, and transaction slippages of different trading pools. This requires a lot of work, and when she signs the transaction, her calculation results may be outdated - because the rates, prices, and slippages are constantly changing.
With intent, the whole process is completely different. Alice only needs to express her intention: I want to buy as many BUZZ as possible with 1,000 USDC on Ethereum, at least 30,000, and the application that handles the intent can complete all the calculations and transactions for her, giving her only the final result: BUZZ on Solana. In short, intention transaction is to build an intent layer on top of the wallet, allowing users to express their intentions and let solvers compete to solve and finally solve the users intentions.
In fact, building the intent layer is very difficult. First, users need a language to express intent. We hope that this language has enough expressive power for any intent, but we also need a protocol to regulate intent and perhaps restrict the language for two reasons: First, unlimited expressions of intent are likely to be unresolved or unverifiable on the chain. Second, expressive power will make intent vague, and vague intent will generate more MEV, which we hope to minimize. This is another difficult balance to achieve.
Anoma, SUAVE, Essential and other projects have tried to build an intent layer. Regardless of the fact that they are still in the conceptual stage, the design often requires the introduction of a domain-specific language (DSL) for intent expression, which greatly reduces the interactive experience for users. Therefore, it is natural to think of using an interactive interface similar to ChatGPT to understand, extract and convert the natural language used by users to express their intent into the standard language of intent, so as to obtain solutions from the solver without perception. The rise of AI Agent has made this idea a reality.
AI Agent, or Artificial Intelligence Agent, is a computer program that can autonomously make decisions and perform tasks based on the environment, inputs, and predefined goals. The core components of AI Agent include a large language model (LLM) as its brain, which enables it to process information, learn from interactions, make decisions, and perform actions; observation and perception mechanisms, which enable it to perceive the environment; reasoning thinking processes, which involve analyzing observations and memory content and considering possible actions; action execution, as an explicit response to thinking and observation; and memory and retrieval, which stores past experiences for learning.
Simply put, AI Agent is an intelligent entity that can act autonomously, perform tasks, and interact with external tools. In the scenario of intentional trading, AI Agent is integrated into the intention layer. After interpreting the intention expressed by the user in natural language, it generates DSL to communicate with the solver, and the process of building the actual transaction is the responsibility of the third-party solver. But in fact, the architecture of intentional trading has changed a lot now. AI Agent has completely replaced the intention layer, and DeFAI has been incubated, which is mainly due to the powerful functions of AI Agent.
First of all, AI Agent can directly act as a solver and improve the efficiency and intelligence level of the traditional problem-solving process, because AI Agent can use its perception, planning, memory, tool use and other capabilities to not only combine smarter trading strategies, but also self-feedback from past trading behaviors to improve its problem-solving capabilities.
Secondly, AI Agent can also integrate modules such as knowledge base, token analysis, and social media publishing. Intention-based transactions mainly focus on account abstraction to help users complete various transaction tasks. AI Agent extends the usage scenarios of intention-based transactions and can solve user needs in all aspects.
Finally, the AI Agent based on the Swarm architecture can achieve mutual collaboration between multiple agents. This collaborative capability enables the system to handle more complex and large-scale tasks, further expand the upper limit of intended transactions, and provide users with complex interaction scenarios.
Function and implementation of DeFAI
Currently, there are DeFAI applications on the market that can realize intention trading, such as Griffain, Hey Anon, Orbit, The Hive, Neur, etc. However, most of them are in internal testing or Beta stage (Griffain, Hey Anon), or not free (Griffain, Neur), or not open source (Hey Anon, Griffain, Orbit). Only The Hive is completely free and open source. Here I would like to express my respect to the author, and then the function introduction will also take The Hive as an example. The application page is https://www.askthehive.ai/chat.
The entire page is dominated by a chat box. LLM supports OpenAI, Anthropic, XAI, Gemini and Deepseek. There are four functional modules below, namely token trends (Trending), staking (Stake), trading (Trade) and knowledge (Knowledge).
The main way to use it is to express your intention through the chat box. For example, if you enter Swap 1000 $USDC to $BUZZ on Solan, The Hive will call the Trading Agent and pop up a plug-in. Click Connect to connect to the (Phantom) wallet to sign and send the transaction.
You can also enter Analysis the trending of $ETH to analyze the recent trend of Ethereum. The Hive will call the Token Analysis Agent to display the basic information of Ethereum, market conditions, price trends, transaction conditions and active wallet conditions, and give trading suggestions.
You can also enter What is DeFi, and The Hive will call the Knowledge Agent to return DeFi knowledge to the user.
The above are just some simple usage examples. The Hive has more functions and is constantly being iterated. For example, the newly added Tokens page can display popular cryptocurrencies and the inflow of smart money, which is helpful for users who are keen on meme, pump.fun and gmgn.
The Hive code repository is https://github.com/ask-the-hive/the-hive. We can briefly analyze the implementation principle.
The Hive defines a list of Agents provided to the outside world. Currently there are 7 Agents:
walletAgent is responsible for the operations of the Solana wallet, encapsulating tools such as querying token balances and transferring tokens;
StakingAgent is responsible for Solana staking related operations, encapsulating tools such as Delegate staking and liquidity staking;
marketAgent is responsible for market-related queries and encapsulates tools such as crypto market conditions, whales, and smart money monitoring;
tradingAgent is responsible for Dex-related operations on Solana, and encapsulates tools such as querying token addresses and Jupiter aggregation;
konwledgeAgent is responsible for knowledge base related queries, and currently mainly supports queries related to the DeFI ecosystem on Solana;
tokenAnalysisAgent is responsible for token-related analysis. Compared with marketAgent, it is more suitable for analyzing specific tokens and also encapsulates tools such as crawling Twitter information.
liquidAgent is responsible for liquidity-related queries of token pairs and encapsulates tools such as Raydium and Dexscreener.
The core logic is to analyze the users input through LLM to find the appropriate Agent. The author has defined the prompt and tool set for each Agent. The AI Agent framework then organizes the Agents workflow and finally returns the execution result to the user. The AI Agent framework here is also implemented by the author himself (the Solana Agent Kit was introduced but not actually used, or was removed in the previous version).
Summarize
Web3 has long been accused of having a poor user experience, hindering the large-scale adoption of innovative applications such as DeFi.
The emergence of “intent transactions” provides a solution, shortening the complex curve of entering DeFi by allowing users to define the on-chain results they want and outsourcing the difficult technical work of achieving these results to third-party solvers (solvers) that interact directly with the network and protocols on behalf of users.
The emergence of DeFAI has revolutionized the curve into a straight line. Using AI technology and providing a natural language interface, users can conduct DeFi transactions and management more intuitively, simplifying the complexity of DeFi operations and improving users on-chain experience, bringing real value to users.
Only when a technology creates real value for users and brings real convenience can it be truly accepted by the public. The emergence of DeFAI is expected to completely change the interaction paradigm of Web3. Even users who are not familiar with blockchain and DeFi can participate in the wave of Web3.
This article was written by pignard.eth (X account @pignard_web3 ) of ZAN Team (X account @zan_team ).
Note: This article is only for technical sharing and does not constitute any recommendation or suggestion.
