Latest Research on Variant: Crypto AI Agent is Becoming a First-Class Citizen in the On-Chain Economy

Author: Mason Nystrom, Partner at Variant

Compiled by: TinTinLand

Bots are becoming "first-class citizens" in the crypto economy.

You can easily observe this trend. Seekers have deployed bots like Jaredfromsubway.eth, leveraging the demand for convenience from real users to execute their trades on decentralized exchanges (DEX) ahead of others. Banana Gun and Maestro allow real users to trade conveniently on Telegram using bots, and on Ethereum, they are among the largest gas consumers. Now, on the new social application Friend.tech, once the platform finds initial real user adoption, bots will join in, potentially inadvertently further driving speculation.

All of this indicates that bots, whether profit-driven (like MEV bots) or consumption-driven (like Telegram bot tool components), are increasingly becoming prioritized users in the blockchain world.

While bots in the crypto space are still relatively primitive, outside of crypto, thanks to the rise of large language models (LLMs), bots have begun to evolve into powerful AI agents designed to ultimately autonomously handle complex tasks and make more informed decisions.

Key Advantages

Building these AI agents in the crypto-native space will bring several significant enhancements:

Native Payment Rails

AI agents can exist outside the crypto space, but if we want AI agents to perform complex operations, they will need to acquire some funding. Compared to allowing AI agents access to bank accounts or payment programs (like Stripe), or dealing with much less efficient matters in the off-chain world, crypto-native payment rails provide a significantly beneficial improvement for AI agents to acquire funds.

AI Agent Wallet Ownership

AI agents connected to wallets will be able to own assets (like NFTs, yields, etc.), granting AI agents the inherent digital property rights of all crypto assets. This is particularly important for transactions between agents.

Verifiable, Deterministic Operations

AI agents will be most effective when operations are provable (in which case they can ensure that certain operations have been completed). On-chain transactions are inherently deterministic—either they have occurred or they have not—meaning that AI agents can complete tasks more accurately than in the off-chain world.

Limitations

Of course, on-chain AI agents also have limitations.

Executing Off-Chain Logic

One limitation is that, for efficient execution, AI agents need to perform off-chain logic. This means that on-chain AI agents will host their logic/computation off-chain to optimize efficiency, but the agents' decisions will still be executed on-chain, allowing for verifiable operations. Importantly, AI agents can also use zkML service providers like Modulus to ensure that their off-chain data inputs are verified.

Dependency on Tools

Another key limitation of AI agents is that their utility depends on the tools provided to them. For example, if you ask an agent to summarize real-time news events, the agent needs to have a web scraper in its toolkit to search the internet for information to perform that given task. If the agent needs to save a webpage response as a PDF, a file system needs to be added to its toolkit. If you want the agent to follow trades of your favorite crypto Twitter KOL, the agent needs access to your wallet and corresponding key signing permissions.

Considering the deterministic to non-deterministic scenarios, most crypto AI agents perform deterministic tasks. That is to say, humans still need to program the parameters of the tasks and how to complete them (like token swaps).

Crypto AI agents have evolved from early keeper bots to today's more complex agents that can leverage LLMs for more complex operations, such as the artist bot Botto that can autonomously create; AI agents that can provide banking services to themselves using Syndicate's trading cloud; and early AI agent service markets from Autonolas.

Cutting-Edge Applications of AI Agents

In cutting-edge fields, various exciting applications have emerged:

"Smart Wallets" Supporting AI Agents

Dawn utilizes DawnAI to provide AI agents that can help users send transactions, execute trades, and perform other real-time on-chain observations (like trending NFTs).

Crypto Game Agents

Parallel Alpha's latest game Colony aims to create AI characters that can own wallets and trade with each other.

Enhanced Toolkits for AI Agents

The practical capabilities of AI agents depend on their toolkits, and interacting with blockchains is currently an emerging field. Crypto AI agents need wallets, funding acquisition, permission functionalities, integrated AI models, and the ability to interact with other agents. More specifically, Gnosis showcases this primitive infrastructure through its AI mechs, where its AI agents encapsulate AI scripts into smart contracts, allowing anyone (including another bot) to call the smart contract to execute agent actions (like betting in prediction markets) while also being able to pay the agent.

Enhanced AI Traders

DeFi super applications provide traders and speculators with advanced capabilities, including: continuously DCAing into positions if conditions are met; executing trades when gas fees drop below a certain price; monitoring new meme token contracts; determining order routing without users needing to know how to onboard, etc.

Long Tail of Building AI Agents

Large AI applications like ChatGPT are suitable for certain general chat scenarios, but AI agents need to be fine-tuned for numerous industries, topics, and niche markets. Markets like Bittensor create incentives for "miners" to train models for specific tasks (like image generation, pre-training, predictive modeling) and train models around target industries (like cryptocurrency, biotechnology, academia). Although Bittensor is still in its early stages, developers have begun using Bittensor to build long-tail applications or agents based on open-source LLMs.

NPC Consumer Application Agents

NPCs (non-player characters) are common in games like MMORPGs, but they are rare in multi-user consumer applications. However, the financialization characteristics of crypto consumer applications make AI agents excellent participants for introducing new gaming mechanics. Open AI infrastructure company Ritual recently released Frenrug, an LLM-based agent that operates in Friend.tech, executing trades (buying or selling keys) based on user messages. Friend.tech users can try to persuade the agent to buy their keys, sell others' keys, or attempt to get the Frenrug agent to use its funds in other ways.

As more applications and protocols begin to use AI agents, people will use them as a bridge into the crypto economy. While today's AI agents may seem like toys, in the future, they will enhance the experiences of everyday consumers and become key stakeholders in protocols, thereby building the entire crypto economy.