AI Agent Article 2: AI Agent Frameworks

The Cambrian explosion of AI agents we've seen in the last couple of months has been enabled to a large part by frameworks that simplify the development of agents.

In the last article of our AI Agent Framework series we've introduced you to the leading agents. Many of these agents were built on exactly such frameworks. In this second article, we examine the leading frameworks enabling this revolution, comparing their approaches and assessing how they tackle the challenges of creating AI agents.

Virtuals

Virtuals is the leading launchpad for AI agents. To complement this platform, they developed the G.A.M.E (General Agentic Modular Engine) framework, a decision-making engine that enables anyone to build AI agents through prompts alone, without writing a single line of code. Advanced developers can also expand upon this no-code foundation by adding custom code to enhance their agents' capabilities.

The G.A.M.E framework's first strength lies in its modular approach to agent capabilities through its unique "locations" system. Instead of making all functions available simultaneously, G.A.M.E organizes capabilities into specialized tool sets that agents can access contextually. The second key innovation is its two-layer decision-making process: The High-Level Planner handles strategic thinking while the Low-Level Planner manages tactical execution, enabling agents to transform high-level goals into concrete actions. 

To integrate agents with social media, G.A.M.E offers two approaches: a plug-and-play option for supported platforms like X, and an API service for custom implementations. Before pivoting to social media agents, Virtuals was initially created with gaming AI in mind. This initial focus is still a strength of G.AM.E and shows today in recent partnership expansions. Virtuals collaborates with Illuvium on creating NPCs with dynamic, intelligent behavior in Illuvium’s upcoming MMO.

Source: https://whitepaper.virtuals.io/

Virtuals’ economic model focuses on integrating their token within the ecosystem. It is required for creating agents, and all agents launched through their platform must use Virtuals as the base token in liquidity pools. For this the developer has to contribute 2400 Virtuals tokens as initial liquidity, locked to be deployed once the agent completes the bonding curve. Further, traders need to use Virtuals to buy freshly released agent tokens until they complete the bonding curve and graduate to Uniswap. Finally, to generate revenues the platform also charges pre and post bonding trading fees of 1%.

This creates a powerful flywheel: as more developers use the platform to launch agents, whether built with G.A.M.E or their proprietary models, more demand is driven to the token and more tokens get locked up, driving more value to the token. This structure has proven successful, positioning Virtuals as a leader in the AI agent space and sending the token to a $5B valuation at its peak.

AI16z

AI16z is built around Eliza, a modular open-source framework for creating AI agents that can interact with both social media and blockchain systems. The project emphasizes open collaboration and community-driven innovation. Eliza's architecture includes a character file system that defines an agent's personality through knowledge, lore, style, topics, and examples, incorporating RAG (Retrieval augmented generation) capabilities for consistent personality across platforms.

The framework offers out-of-the-box social media integrations across Discord, X, and Telegram, a trust engine for autonomous blockchain interactions with built-in risk management and validation checks and an extensive plugin system that enables continuous community-driven expansion of agent capabilities. Through these plugins, developers can add features like image generation, blockchain integration, and secure execution environments.

AI16z focuses on developing agent swarms - networks of AI agents working collaboratively with humans and each other. This vision becomes reality through Eliza V2's scalable message bus and improved concurrency management, enabling complex agent-to-agent interactions at scale. To accelerate progress, AI16z has secured a partnership with Stanford University, where researchers are developing trust mechanisms for AI agents and investigating how agents can coordinate with and govern each other, complementing AI16z's technical foundation for agent swarms.

Value accrual for the AI16z token currently happens through voluntary token contributions from projects to the DAO. Further, projects launching through Vvaifu (a popular community launchpad) pay deployment fees in both SOL and token supply. There's speculation about an official ELIZA launchpad that could enforce token contributions through smart contracts. However, as an open-source framework, ELIZA will always remain freely available for independent use outside these launchpads. Finally, the token also represents a DAO investment vehicle set to distribute all investments and profits to holders in October 2025.

The ecosystem's rapid growth is evident in its GitHub metrics, with over 6,100 stars and 1,800 forks from 193 contributors Eliza has become the Repo with the most stars on Github. A Creator Fund, established through a significant token holder's contribution, further supports developer engagement.

Source: star-history

Zerepy 

Zerepy is a python based framework that powers Zerebro, the creative agent we covered in our first article of this series. Initially developed just for Zerebro, Zerepy is now publicly released to enable other developers to create their own agents with this framework. 

Zerebro’s soon to be open-sourced LLM removes corporate guardrails to enhance creativity and adaptability, aiming to create a system capable of human-like innovation and learning. On top of their own model, Zerepy supports multiple LLM providers such as OpenAI, Anthropic, EternalAI and integrates with diverse social platforms like X and Farcaster.

While initially focused on creative and social media capabilities, Zerepy is rapidly expanding its scope. The framework provides sophisticated vector memory systems for data processing, crypto-native features for on-chain activities, and cross-chain capabilities through LayerZero integration. Through partnerships like askthehive.ai, ZerePy is developing agent-to-agent communication protocols that will enable agents to coordinate on-chain.

All these components come together in Zentients.xyz, their upcoming consumer-facing agent platform. It will serve as a launchpad incorporating multi-agent intelligence, decentralized GPU renting, and cross-chain interoperability, positioning ZerePy as an L1-like foundational architecture that allows agents to operate across any stack, chain, or token.

The economic model of this platform ties to the $ZEREBRO token. The framework's expansion into DeFi and the launch of Zentients creates multiple value accrual mechanisms: launch fees, LP agent pool fees, taxes on fees agents generate, and potentially GPU rental fees. This structure aims to create a sustainable ecosystem where increased agent deployment and usage drives value back to the token while funding continued development.

Arc

The platform arcdotfun and the RIG (Rust Infrastructure for Games) Framework are developed by Playgrounds. Unlike many AI frameworks that use Python or TypeScript, RIG is built on Rust, making it a purpose-built system for developing performant, secure, and modular AI agents.

Rust provides superior performance and memory safety, with strict compile-time checks and zero-cost abstractions. These technical advantages become particularly important when dealing with complex AI systems where multiple agents interact with each other.

However, this technical sophistication comes with trade-offs. The primary disadvantage is that development with Rust is more challenging and the pool of available developers is smaller, as Rust has a steeper learning curve compared to Python or TypeScript. This could potentially slow down ecosystem growth in the short term.

Yet this selectivity aligns perfectly with arc's strategy, as demonstrated by their Handshake program. While anyone can build using the RIG framework, only projects that pass the team's vetting process can become part of this program allowing them to launch on arcdotfun. 

Projects participating in the Handshake program receive official endorsement from Arc and will be reviewed by the Solana Foundation and Arbitrum DAO for grants funding. This filter and Rust's technical barrier should ensure that while RIG might see fewer developers overall, the agents launched on arcdotfun have a higher quality.

Agents launched on arcdotfun commit to sharing a portion of their generated fees with Arc. Those revenues are used for continued development of the RIG framework. The token alsoserves other key functions in the ecosystem. When new agents launch on arcdotfun, they can pair their token with ARC in liquidity pools. Though unlike Virtuals, Arc doesn't seem to force that and might allow other base pairs. Token holders can stake ARC to participate in governance decisions, and holding ARC is required for accessing proprietary AI models within the ecosystem.

This structure aims to create a flywheel: as the framework improves through fee-funded development, it attracts more high-quality developers and projects. These new projects generate fees and pay tax revenue to Arc, accelerating framework development further and strengthening the entire ecosystem eventually attracting even more quality developers.

The Rei Framework

The REI Framework is what powers Rei_00, the agent providing the most advanced financial analysis we introduced in the last article of this series. It enables AI agents to store their experience and knowledge directly on the blockchain to enable a transparent learning process.

This specific combination of AI and blockchain tech is challenging, as AI systems given the same input might produce different outputs each time as they learn and adapt. Blockchains systems on the other hand need exact consistency. Rei’s innovative system bridges the gap between this probabilistic AI thinking and deterministic blockchain execution through a sophisticated four-layer architecture and specialized data standards.

The system relies on three key components: The Oracle Bridge converts varying AI outputs into consistent, verifiable results, the ERCData standard enables storage of complex relationships while maintaining determinism, and the Memory Systems allow agents to build upon past experiences stored on-chain. These components work together with a four-layer processing system: Thinking (pattern recognition), Reasoning (context addition), Decision (action determination), and Acting (blockchain execution).

Source: 0xreisearch.gitbook.io

For AI agent developers, REI offers a unique value proposition: the ability to create self-improving agents whose increasing knowledge and decisions are permanently stored on-chain, making them fully transparent and verifiable. The framework includes development tools for memory management, state verification, and efficient on-chain storage, with reasonable operational costs thanks to its EVM native architecture.

The economic model currently focuses on their oracle network implementation, utilizing the $REI token in a tiered fee structure. Computing providers, validators, and the protocol treasury receive portions of these fees, with staking and slashing mechanisms to ensure network reliability. However, REI's broader token flywheel and value accrual mechanisms beyond the oracle fees remain unclear.

Comparative Dimensions

Primary Focus and Target Market 

• Virtuals aims for mass adoption through maximum accessibility, prioritizing rapid growth by enabling non-technical users to create AI agents. Their innovation lies in combining no-code simplicity with sophisticated agent behavior, initially targeting gaming but expanding into broader content creation where contextual adaptation is crucial. Therefore their G.A.M.E. framework is best suited for more simple twitter agents that can be set up quickly.

• Eliza drives community-driven agent development, emphasizing open-source collaboration and the creation of interconnected agent networks. Their swarm-focused vision is strengthened by Stanford's research partnership, developing advanced trust mechanisms and coordination protocols to enable effective agent-to-agent interactions. This makes Eliza the best suited framework for agents that should handle complex tasks and need blockchain integration.

• ZerePy revolutionizes creative AI by removing traditional constraints, enabling agents to generate diverse media outputs from text to video. They complement this creative focus with robust infrastructure development, combining cross-chain capabilities through LayerZero with dedicated GPU resources to support their agents' computational needs. Zerepy is therefore best suited for agents that should generate creative output and monetize it through various blockchains.

• RIG prioritizes enterprise-grade performance and reliability, carefully curating their ecosystem through their Handshake program. The program offers benefits to selected projects, including official endorsements and connections to get access to grant funding. RIG is the best framework for complex multi-agent systems that require a lot of computation and rely on superior performance and security. It’s the best framework for enterprise AI agent solutions.

• REI enables true learning and evolution of agents through on-chain memory and experience storage, making their decision-making process transparent and verifiable while maintaining flexibility in agent. This makes REI the ideal framework for use-cases where gradual learning and adaptation is crucial, such as quantitative analysis and trading.

Technical Foundation and Language Choice 

• G.A.M.E combines easy accessibility through prompt-based development with customization options for more advanced developers through its language-agnostic API and SDK system. Technical innovation lies in two systems: A unique architecture that organizes agent capabilities into contextual toolsets, and a two-layer planning system that enables agents to transform abstract goals into concrete actions.

• Eliza builds on TypeScript to balance accessibility with robust functionality. Their extensive plugin system enables easy expansion of agent capabilities. For multi-agent coordination, Eliza V2 introduces a scalable message bus and improved concurrency management, providing the technical foundation for complex agent-to-agent interactions.

• ZerePy leverages Python's rich AI ecosystem while building multiple technical pillars: their own creative LLM removes traditional guardrails for dynamic behavior, LayerZero integration enables cross-chain operations, and dedicated GPU infrastructure supports computational needs. It also offers agent-to-agent communication protocols.

• RIG's implementation in Rust ensures high performance and safety through zero-cost abstractions. The framework includes a flexible provider layer for multiple LLMs and vector stores, making it particularly adaptable for data-intensive enterprise applications.The performance and security benefits Rust also enable safer and more efficient agent-to-agent communication.

• REI's innovative architecture bridges the fundamental gap between probabilistic AI thinking and deterministic blockchain requirements. This innovative approach enables agents to learn and evolve while maintaining blockchain compatibility, opening possibilities for agents to learn any task while remaining verifiable.

Economic Models and Growth Strategy 

• Virtuals was the first team to establish a value accrual flywheel by requiring their token for agent creation and mandating it as the base pair in liquidity pools. New agents must initially trade against Virtuals tokens, creating consistent demand and token lock-up.

• Eliza combines voluntary token contributions (1-10% from projects) with deployment fees (1.5 SOL + 5% token supply) on their coming community launchpads. Their DAO investment vehicle structure adds another avenue to value creation.

• RIG generates platform revenue through their quality-focused approach, where high quality agents share operational revenue with the platform. With these fees funding continued framework development and attracting more quality projects. Further access to their proprietary AI models requires token holdings, boosting token demand.

• ZerePy's Zentients platform will capture value through multiple streams: launch fees, LP agent pool fees, taxes on agent-generated revenue, and GPU rental fees, creating a comprehensive value accrual system.

• REI currently focuses on oracle network fees distributed between computing providers, validators, and treasury. Further having access to their framework will require REI tokens, though it is not yet specified how exactly the token plays into access.

Strengths and Challenges 

• Virtuals's established ecosystem combines the strength of low-code integration for broad accessibility with a proven token model and strong network effects. However, the limited scope of the G.A.M.E framework might make them vulnerable to more technically advanced frameworks - offering more sophisticated capabilities - eating into their market share. 

• Eliza's strength lies in its community-driven development and technical innovation in agent swarms. Their extensive plugin system enables rapid development of diverse agents while the framework grows more versatile as the community contributes. Stanford's research partnership reinforces their leadership in multi-agent coordination. However, their open-source community driven approach might slow down progress as coordination is harder.

• ZerePy leverages Python's widespread adoption in the AI community while offering unique value through their unrestricted creative LLM. Their multi-chain integration, GPU power solutions and cult following provide strong growth potential. Though their limited scope on creative output might cap the size of their targetable market.

• RIG's modular architecture and Rust foundation delivering superior performance and reliability is attractive to enterprises who require these capabilities and have access to top tier developers. However, Rust's steep learning curve and Arc’s selective approach create a high entry barrier limiting community growth. This slows down their ecosystem growth, which is risky in such a fast moving space.

• REI's breakthrough in blockchain-native learning represents a fundamental shift in how AI agents can evolve - enabling them to learn from experience and improve themselves over time. While this positions them uniquely in the market, they face the challenges of entering a market with already well established players.

Future Outlook

The rapid evolution of AI agents in recent months has been accelerated by frameworks that provide developers with powerful tools to experiment and innovate. From AI16z's vision of interconnected agent swarms to REI's breakthrough in on-chain learning, these tools are enabling developers to create agents that interact, learn, and evolve autonomously in ways previously impossible.

The technical designs of these frameworks remain experimental, with each framework taking radically different approaches to agent development. This diversity of solutions shows just how early we are in this space. While the frameworks currently serve different niches, the rapid pace of innovation suggests these market segments aren't yet settled. We're still discovering the fundamental building blocks of AI agent development. At this stage, it's too early to determine whether these distinct approaches will evolve into established solutions within their respective niches, or whether a few approaches will emerge as clear winners by achieving superior product-market fit.

On the economic front, we've already seen such convergence toward a winning approach. Virtuals proved that a comprehensive token model built around a launchpad including demand driving mechanisms is highly successful. Other frameworks are now following this blueprint, developing their own launchpads and implementing similar token flywheel mechanisms.

This foundation of technical innovation and proven economic models sets the stage for our next article, where we'll explore why the convergence of AI agents and crypto could represent one of the most significant developments in both spaces. We'll evaluate the potential of this new AI x Crypto space and analyze how this merger might reshape our understanding of both artificial intelligence and blockchain technology. Stay tuned as we delve into why the AI agent space might become the defining narrative of the next crypto cycle.