Research: AI Agent Sector Overview

Early AI agent implementations in the crypto space primarily leveraged automated social chat functions, sentiment analysis and surface-level on-chain analysis.

These initial developments, referred to here as the “first wave,” laid the groundwork for conversational AI within platforms such as X, enabling tasks from content moderation, user engagement and on-chain data retrieval.

The sector then matured as the “second wave” of AI solutions began to emerge to include complex DeFi protocols, cross-chain interoperability and smart contract–based automation.

It is worth noting that while these projects speak of “AI-driven” operations, the heavy lifting of inference is conducted off-chain.

This report aims to offer an overview of these projects while evaluating how they collectively contribute to the sector’s future.

Individual Project Analyses

Agentic Frameworks

The first wave of AI agents can be illustrated through various projects that focused on agentic frameworks, from ai16z & its ELIZA framework, Virtuals and its G.A.M.E. framework, the AI RIG Complex (ARC) framework, and Coinbase’s AgentKit, all of which gave rise to mostly, but not exclusively, AI agents that emphasized communicative functions.

Originally positioned as a DAO-oriented initiative aimed at integrating AI into fund management, ai16z’s limited transparency of its in-house AI agent “AI Marc” and the investment fund it supposedly managed tempered its effectiveness.

ai16z’s true traction emerged from its ELIZA framework, a TypeScript-based toolkit for AI agent development across platforms such as X, Discord and Telegram, among others, as well as on-chain environments such as Solana or EVM-based blockchains.

Although ELIZA’s core AI processes run off-chain, with only essential outputs transmitted to blockchain systems, its design demonstrated early integration of AI insights with on-chain activities.

The framework aims to let anyone develop agents with varying functionalities, from basic chatbots or plugins to potentially advanced functionalities such as integrations that tie an ELIZA agent into proprietary market data services or specialized analytics dashboards.

An in-depth report on ai16z and the ELIZA framework was conducted in a previous research piece .

The modular nature of the ELIZA framework gained traction among developers, with nearly 15K stars on Github at the time of writing, while its related $ai16z token reached a market cap of over $2.5B at its peak, though it has since declined by over 80% at time of writing.

GitHub activity, judging from the number of stars the elizaOS repository receives per day, has also declined significantly since its peak in December 2024.

Source: sentient.market, The Block Pro Research

On the other hand, Virtuals is essentially a launchpad for AI agents, while its G.A.M.E. framework is similar to ELIZA where it enables an agent to plan and execute actions and decisions based on the information provided to it.

An overview report on Virtuals was conducted in a previous research piece .

The most notable AI agent that came from the Virtuals ecosystem was aixbt, with the ability to autonomously analyze on-chain and off-chain crypto market data to provide real-time insights and interact with users on X.

ARC was another agentic framework with similar intents and purposes as both ELIZA and G.A.M.E., with its main differentiator being based in the Rust language.

Similar to the ELIZA framework, ARC’s GitHub activity, judged from the number of stars its repository receives per day, has also declined significantly since its peak.

Source: sentient.market, The Block Pro Research

On the other hand, Coinbase’s AgentKit is a model-agnostic framework that integrates directly with the Coinbase Developer Platform SDK and supporting tools like LangChain that enables developers to create, customize and maintain blockchain-aware agents.

DeFAI and Transactional AI Solutions

In the “second wave”, projects such as HeyAnon, Wayfinder, AgentKit, Giza and its ARMA agent, as well as Almanak showcase a growing emphasis on automated transactional flows and decentralized finance AI (DeFAI). 

Unlike the first wave, these projects are designed to work within established DeFi frameworks, where AI models operate off-chain and transmit signals for on-chain execution.

This approach acknowledges that while smart contracts remain immutable and deterministic, they can benefit from off-chain AI that analyzes data and suggests actions, provided these outputs are securely and reliably integrated.

It is worth noting that the vast majority of projects in this category have yet to launch tokens of their own.

An exception to this is HeyAnon, which integrates natural language processing models with DeFi protocols, allowing users to execute swaps, lend assets or bridge tokens by submitting text-based instructions through a conversational interface.

Implementing a similar concept but without a token as of yet, Wayfinder leverages AI agents called “shells” that navigate DeFi protocols through a chat-based interface.

By using “wayfinding paths”, essentially curated routes across different blockchains, Wayfinder lets users perform DeFi tasks by typing out their intent in everyday language.

Although the vision of a single, AI-driven “super app” for DeFi is compelling, questions remain regarding data reliability and safeguarding against malicious smart contract triggers, even despite the inclusion of built-in risk mitigations, such as fallback procedures if the AI output deviates significantly from the expected output.

It is also worth noting that this system relies on LLMs to interpret smart contract logic off-chain before their conclusions are fed into on-chain processes.

With this in mind, potential misinterpretation and the possibility of malicious or flawed paths are significant inherent risks.

Giza is an AI agent stack that offers frameworks, command-line interfaces, datasets and SDKs for autonomous agents.

Its debut AI agent, ARMA, optimizes stablecoin yields across multiple DeFi lending protocols by dynamically reallocating funds and auto-compounding returns. 

ARMA relies on real-time APR data and frequent rebalancing, presumably handled off-chain with only secure summaries fed on-chain, highlighting the importance of reliable data inputs and minimizing computational overhead.

While early data suggests potential yield improvements, its long-term success will depend on robust security audits and performance under market volatility.

Almanak provides a comprehensive suite for AI-driven DeFi strategy ideation, testing and deployment, made up of two main components: Its strategy and agentic infrastructure.

The “strategy infrastructure” uses a blockchain state machine for simulation, offering more realistic modeling than traditional price feed–based approaches. 

Meanwhile, its “agentic infrastructure” leverages LLMs such as Llama and Mistral to automate the full lifecycle from strategy ideation to on-chain execution.

Techniques such as privacy-preserving execution via trusted execution environments are utilized to help mitigate front-running risks.

Story stands apart from the wave of DeFAI-driven agent deployments by focusing on on-chain intellectual property (IP) rather than direct AI-agent execution.

Story tokenizes IP into programmable assets, enabling automatic licensing and royalty distribution through specialized “cores” on a custom Layer-1 architecture.

Its Agent Transaction Control Protocol (Agent TCP/IP) aims to facilitate AI-driven licensing negotiations for IP assets.

An in-depth report on Story was conducted in an earlier research piece .

However, it is worth noting that Story is a heavily–funded project with significant private market raises, hence resulting in its comparatively higher valuation in the public market at launch.

Sector-wide Meta Analysis

The development trajectory observed across these projects indicates a clear shift from social facilitation to direct financial engagement.

The first wave demonstrated that AI had the potential to be an effective tool for orchestrating chatbots, moderating online communities and even proposing governance ideas for DAOs in an advisory capacity, though the actual execution surrounding the latter was fairly lackluster in retrospect.

Projects in this category concentrated on building specialized frameworks for personality-rich bots and bridging AI’s language capabilities into blockchain-based interactions.

The success of those earlier ventures, however, also revealed gaps in real economic application.

As a result, the second wave builds on these early experiments by targeting transactional contexts, bridging multiple chains and automating DeFi operations at scale, with projects such as HeyAnon and Wayfinder, as well as Giza, ARMA and Almanak, exemplifying this shift.

While each of these second-wave projects approaches transactional engagement differently, they all illustrate a broader movement away from AI as a mere “chatty bot” and toward AI as an integral participant in value exchange.

This shift brings deeper challenges and tensions, raising questions on how these projects should handle potential exploit risks, whether from sophisticated market manipulation, AI misinterpretation, data signal corruption or vulnerabilities in off-chain processing pipelines.

On the other hand, Story underscores the expanding definition of AI facilitation by applying programmable on-chain logic to IP licensing and usage scenarios, including AI data sets.

Limitations

Several limitations become evident across these platforms:

Security Architecture

Much like every protocol or mechanism in crypto, AI-driven on-chain interactions have inherent risks of private-key compromises and the exploitation of vulnerabilities in real-time rebalancing or licensing protocols.

Data Reliability & AI Model Risks

Unverified or manipulated data sources can undermine yield-optimization or attention-token strategies. LLMs also remain susceptible to errors and “hallucinations.”

Robust guardrails, such as rigorous model validation, continuous monitoring and using consensus from multiple AI systems, would be essential to prevent misinterpretation of critical instructions.

Social Automation Challenges

AI-driven social automation relies on processing vast amounts of data off-chain and then feeding actionable insights on-chain.

Given that the bulk of data analysis happens off-chain, ensuring that these systems remain transparent and verifiable is a significant challenge.

Developers must consider not only the technical aspects of data processing but also how to implement oversight mechanisms that ensure fairness and mitigate bias, particularly when AI is potentially used for governance processes.

Takeaways

The first wave of AI projects in the blockchain space illustrated that automated chat tools, social media bots and community-driven interactions have a place to thrive in. Yet these initiatives also exposed the boundaries of conversational AI, highlighting the ceiling in the lack of economic utility.

The second wave has taken a decisive step forward by integrating AI into DeFi protocols and advanced on-chain operations.

Despite the promise, DeFAI may also inherit the private-key vulnerability profile of on-chain trading bots.

Similar to trading bots, if the UX gains of DeFAI prove compelling, mainstream adoption could increase exponentially regardless. However, a portion of the user base will inevitably suffer from the tail risks of key mismanagement or malicious exploits, if they were to occur.

In short, the journey from chat-based novelty to value-based practicality is now well underway. If these obstacles are tackled with thorough testing and fail-safe mechanisms, AI could indeed become a pivotal enabler of user-friendly DeFAI.