AI Agents 2024 Review and Outlook: Where Have We Come From and Where Are We Going?

The original text is from 0x Jeff

Compiled | Odaily Metaverse Report Golem（@web3_golem）

Review of AI Agent Development in 2024

2024 was a year of transformation for AI Agents. About three months ago, truth terminal attracted everyone's attention with its witty personality, conversational style, and interaction with A16z co-founder Marc Andreessen, becoming the first "millionaire agent" and triggering a trend of AI Agents.

Soon after, Virtuals also entered this field, pioneering "Agent Tokenization" and consolidating this narrative. Since then, innovation has been explosive:

• Luna: This agent launched a on-chain wallet tipping feature for fans, and can now browse Twitter, analyze posts, and even join Google Meet.

• Conversation agents on Twitter/X: Some agents have become "meme masters", while others focus on acquiring and sharing Alpha information. For example:

1. aixbt: Known for its concise, actionable insights and some light-hearted trolling;

2. Dolos: Sharp-tongued, has now developed its own framework and provides support to other agents through Dolion.

Meanwhile, AI Agents are becoming more entertaining, with 3D models, voice capabilities, and cross-platform presence. Representative agents include:

• AVA and Holoworld AI: The first 3D audio-visual framework, endowing agents with 3D bodies, voices, and deeper personalities;

• zerebro: A music agent that has released top-tier albums and is about to launch its own framework ZerePy, allowing more people to build agents like Zerebro;

• Nebula: A Meme AI KOL that can create meme images, videos, and appear in AR/VR environments and games;

• LucyAI: The first realistic anime agent that can speak multiple languages, live stream, and interact with fans;

• DO KWEEN: A cinematic agent that produces Netflix-quality episodic content weekly.

AI Agent Narratives in 2024

Meanwhile, the a16z and open-source innovation movement have also gained attention, especially after the launch of the Eliza framework. Developers have come together to create toolkits, plugins, and other functionalities, fostering collaboration and innovation. During this period, Virtuals has grown into a unicorn company, further consolidating its leadership in the AI Agent distribution platform.

The open-source innovation movement has sparked the interest of the developer community, kicking off the biggest community collaboration of the year. Currently, more and more projects are emphasizing the importance of "open-source frameworks". As agents continue to evolve, new narratives have emerged to promote more agent collaboration:

• Agent Metaverse: Proposed by Realis, it created a Minecraft map version of a replica Earth to house these AI agents, allowing them to interact and build a civilization.

• Gamification of Agents: ARC Agents is a representative in this field, combining AI with reinforcement learning games. By combining AI and reinforcement learning games, they have launched games similar to Flappy Bird, where agents compete against each other, and the community contributes game data to help these agents grow. ARC recently revealed its vision of moving towards AGI.

• Clustering/Collective Intelligence: FXN is a representative in this field, aiming to establish a unified economy for AI agents, with the idea of clustering being to allow AI agents to work together to achieve common goals. Virtuals is also promoting interaction (or commercialization) between agents, which is a communication protocol that allows agents to seamlessly provide services to each other. Meanwhile, Story has also announced a protocol for agent-to-agent communication for IP, allowing agents to tokenize, monetize, buy and sell IP.

Alongside these narratives, we can also see:

• On-chain Trading Agents: Initially proposed by Spectral, their Syntax v2 allows users to launch trading agents that can trade on Hyperliquid. They have been dominating this field, but progress has been temporarily halted due to a small vulnerability. Another noteworthy agent is Big Tony, which uses Allora's machine learning price prediction model to automatically trade mainstream currencies.

• InvestmentDAO: Initially represented by a16z, but now more DAOs are emerging, such as AIrthur Hayes and Aimonica. The general narrative is that these DAOs raise SOL on daos.fun (or other platforms) and use these funds for trading and investment to generate profits. Creating an InvestmentDAO with the name of a Crypto VC or celebrity makes it more attractive.

• DeFi Agents: Led by Mode, it is the preferred ecosystem for DeFi agents. The main application scenarios include AI-driven stablecoin mining, liquidity provision, lending, etc. There are high-quality teams within the ecosystem, including Giza, Olas, Brian, Sturdy, QuillAI Network, etc.

• AI App Store: ALCHEMIST AI is a representative in this field, providing a no-code tool that allows users to create applications. MyShell is another AI App platform with a larger developer and user base, especially in the Web2 domain.

• Abstraction Layer: griffain and Orbit are representatives in this field, providing a chain-based abstraction experience for all on-chain content, making it easier for users to operate on-chain, especially user-friendly for ordinary users.

• On-chain VC Agents: sekoia virtuals aims to become the primary "rubber stamp" for quality agent projects, currently strictly selecting and investing in only three projects, setting a precedent for on-chain VC.

Data and Framework Development

• Cookie DAO is becoming the main source of AI proxy data and social metrics, with industry insiders relying on it to track proxy influence, market value, and performance;

• Masaintegrated with Virtuals, providing proxies with real-time data,enabling them to self-learn and self-improve;

• TAOCAT is the first proxy driven by the Bittensor subnet, showcasing the possibilities of real-time data (it is the only proxy token that has skyrocketed while other proxy tokens have declined);

• AgentTankshowcased a framework where they brought proxies onto computers, giving them full computer operability, so that they can engage in entertaining interactions and provide interesting commentary on the internet.

Other new frameworks:

• arc: a Rust-based RIG framework that has gained attention for its versatility;

• Dolion: evolved from Dolos, becoming a toolbox for creating unique proxies.

What Did We Learn from 2024?

The above may have overlooked some minor narratives or AI Agents, but through the development of AI Agents, we can learn the following from this year:

Top Teams Valued Over $50M All Have Their Own Tuned Models

They initially showcase the use cases and uniqueness of their proxies, then provide no-code frameworks for others to have proxies as good as their flagship ones, which also helps drive higher proxy value and proxy token prices.

But this does not mean you should build your own framework, or not build on top of any other frameworks like Virtuals G.A.M.E and ai16z Eliza. If you don't have enough AI resources or capabilities, you should join these communities, as the tools can help you quickly realize your ideas and experiment. At the same time, you should leverage Virtuals and ai16z for distribution/marketing, as these two currently provide the best visibility, and integrating and collaborating with them is definitely positive EV.

Investing in Proxy with Own Frameworks or the Entire AI Proxy Ecosystem Will Yield Better Risk-Adjusted Returns

If they manage to create a framework that people are willing to pay to build proxies on, it means that framework has enough attention and demand to drive or sustain the price. arc is a good example, the first Rust framework quickly gained popularity and saw its price rise accordingly.

On-Chain and DeFi Applications Will Be the Product-Market Fit (PMF) for Crypto AI

I believe the areas currently bringing the most value include:

• Abstraction layers to help people navigate on-chain;

• Alpha proxies sharing quality alpha that people can make money from;

• Execution proxies that can help simplify trading, mining, providing liquidity, lending execution.

Perhaps we'll soon see a proxy that combines alpha discovery + execution.

Data is an Indispensable Part of Every Proxy

Bad data = bad output. If data is the new gold, then platforms like Cookie DAO are essentially gold mines. vana is an interesting L1 that tokenizes data into data liquidity pools, they have a DataDAO model that can help people collectively own data, bring in data, clean it for AI Agents, although the token economics may have issues, the product is very interesting.

Looking Ahead to AI Agent Development in 2025

Above, we discussed the development of AI Agents in 2024, reviewing the milestones and innovations of that year. Now, let's look ahead to 2025 - I believe this year will not only see AI Agents become more useful, but also start to reshape our perceptions of autonomy, intelligence, and collaboration.

Setting the Stage for 2025

Before moving forward, it's worth emphasizing that Virtuals will continue to consolidate its position as the primary distribution network for AI Agents on Base. Virtuals has become the go-to platform for agents to pair their liquidity, increase visibility, and form deeper partnerships with other quality projects. The total market cap of proxies on Virtuals is currently around $3 billion, accounting for 77% of the entire AI Agent space (data from Cookie DAO).

As more unique proxies emerge on Virtuals, and as these use cases become increasingly diversified, more developers will be attracted to the Virtuals platform, whether or not they already have a token. This growth will also drive the appreciation of the VIRTUAL token.

While the ai16z Dao has led the open-source innovation movement with its Eliza framework, it currently lacks a launch platform, and the value accrual of its token economics is not as strong as Virtuals. However, the future still holds potential. They have recently formed a working group to improve their token economics, and their future launch platform may propel ai16z to become the top distribution platform on Solana, surpassing the existing launch platforms (if they decide to launch one).

In 2025, we will also see top proxies with product-market fit (PMF) receive major capability upgrades. For example, AIXBT has already taken a leading position in the alpha-focused conversational proxy space and may further consolidate its position through sharper responsiveness and more insightful analysis.

As other vertical leaders emerge, this evolution will be reflected across the ecosystem, with them leading the trends through their unique expertise and innovations.

What Will the Trends Be in 2025?

2025 will be the year of AI Agent specialization. We will see the emergence of vertical leaders in various domains, each dominating their own niche:

• 3D Models: Proxies with high-quality visual design, suitable for gaming, AR/VR, etc.

• Voice Modules: Proxies that can speak naturally, like humans, and evoke emotional resonance.

• Personality-Rich Proxies: Proxies with unique, relatable personalities.

• Live Streaming Proxies: Interactive proxies thriving on platforms like Twitter/X and YouTube.

• Automated Trading Proxies: Capable of executing profitable trades consistently.

Just as humans are diverse and specialized, AI Agents will also become equally diverse. The uniqueness of each agent will be closely tied to its underlying model, data, and infrastructure. However, the success of this ecosystem depends on a robust decentralized AI infrastructure.

The Role of Decentralized AI Infrastructure

Decentralized infrastructure is crucial to scaling AI agents by 2025, as without it, the field risks bottlenecks in performance, transparency, and innovation.

Here's why each component of decentralized AI infrastructure is important, and the projects currently being built to address these challenges:

Verifiability

Trust is the foundation of decentralized AI. As AI Agents become more autonomous, we need systems that allow us to verify what's happening behind the scenes. We need to know whether this "agent" is a genuine AI or a human masquerading as one; whether the output results are accurate and generated by the claimed algorithm or model; whether the computation is executed correctly and securely, and so on.

This also involves Trusted Execution Environments (TEEs), which ensure the agent runs independently, securely, and without manipulation. Without verifiability, there is no trust, and without trust, the ecosystem cannot scale.

Notable projects:

• ORA: Exploring infrastructure for secure AI, but token economics still need refinement;

• Hyperbolic: Pioneered sampling proofs for verifying AI computation and reasoning;

• Phala Network: Known for its TEE infrastructure, adding a layer of security for decentralized AI.

Payments

For AI Agents to operate autonomously in the real world, they need payment systems. Whether transacting with humans or other agents, these systems must handle everything from on-/off-chain to barter and accounting. Imagine agents independently managing finances, purchasing compute resources, or even exchanging services with other agents - this is the foundation for agent-to-agent commercial transactions.

Notable protocols:

• Crossmint: AI payment tools to facilitate transactions;

• Nevermined: Supports commercial transactions and interactions between agents;

• Skyfire: Focused on payments and accounting for agent operations.

Decentralized Computing

AI's compute demands are skyrocketing - doubling about every 100 days. Traditional cloud services like AWS cannot meet this demand in terms of both cost and accessibility. Decentralized compute networks allow anyone with idle resources to join the network, provide their compute power, and earn rewards.

This year, we've even seen the emergence of GPU-backed debt financing models, like GAIB, helping data centers finance and scale their operations. This makes decentralized computing accessible to a wider audience.

Notable protocols:

• Aethir: Decentralized computing tailored for AI and Web3.

• io.net: Scalable compute solutions for AI workloads.

Data

If AI is the brain, then data is the oxygen. The quality, reliability, and completeness of data directly impact the performance of AI models. But the cost of acquiring and labeling high-quality data is high, and poor data leads to poor results.

Excitingly, there are platforms emerging that allow users to own their data and monetize it. For example, vana allows contributors to tokenize their data and trade it in a Data Liquidity Pool (DLP). Imagine choosing the TikTok DataDAO or Reddit DataDAO to pool your contributions - this concept empowers users while driving AI progress.

Notable protocols:

• Cookie DAO: A reliable source of data metrics and insights;

• vana: Tokenizes user data into a tradable liquidity pool on a decentralized marketplace;

• Masa: Collaborating with Virtuals to build the largest decentralized AI data network, supporting dynamic and adaptive AI agents.

Model Creators and Marketplaces

2025 will see an explosion of new AI Agents, many of which will be supported by decentralized models. These models will be more advanced, incorporating human-like reasoning, memory, and even cost-awareness.

For example, Nous Research is exploring a "hunger" mechanism that introduces economic constraints to AI models. If an agent cannot afford the cost of reasoning, it will effectively "die," teaching it to prioritize tasks more efficiently.

Notable projects:

• Nous Research: Introducing a "hunger" mechanism to teach AI resource management;

• Pond: Collaborating with Virtuals to provide tools for decentralized model creation and training;

• Bagel: Providing privacy-preserving infrastructure using FHE and TEE.

Distributed Training and Federated Learning

As AI models become larger and more complex, centralized training systems will be unable to meet the demand. Distributed training spreads the workload across multiple decentralized nodes, making the process faster and more efficient. Meanwhile, federated learning allows organizations to collaborate in training models without sharing raw data, addressing the key privacy concerns.

FLock.io is the "Uber for AI." Flock connects AI engineers, model proposers, and data providers to create a marketplace where AI models can be trained, validated, and deployed in a secure, decentralized manner. It supports projects like Aimonica and many other exciting models.

Collective Intelligence and Coordination Layer

As more specialized agents enter the ecosystem, seamless communication between them becomes crucial. Collective intelligence allows agents to work as a team, pooling their capabilities to achieve common goals. The coordination layer abstracts the complexity, making it easier for agents to collaborate.

For example, Theoriq uses meta-agents to identify the most suitable agents for a task and form "collectives" to achieve objectives. It also tracks reputation and contributions, ensuring quality and accountability.

Notable projects:

• FXN: Creating a unified protocol for communication and commerce;

• Virtuals: Enabling agent-to-agent interaction and integration;

• Theoriq: Developing agents and advanced coordination tools, including clustering and task delegation, for AI agents.

Why Decentralized Infrastructure is Crucial

The next stage of AI agent development depends on infrastructure. Without verifiability, payment systems, scalable computing, and robust data pipelines, the entire ecosystem risks stagnation. Decentralized infrastructure helps solve these issues by providing trust and transparency, scalability, collaboration, and empowerment.

Of course, there are several other narratives that are expected to develop by 2025, such as:

• Proxy Metaverse / AI x Gaming: Projects like Realis and ARC Agents are integrating proxies with games and immersive virtual worlds;

• On-chain and DeFi tools: Protocols like Almanak, Wayfinder, Axal, Cod 3 x, griffain, and Orbit are building fundamental tools for DeFi-driven proxies.

Summary

2025 will be the year of AI Agents, when we will see them rapidly move towards sentient AGI. These agents will not only execute isolated tasks - they will also trade autonomously, collaborate with other agents, and interact with humans in ways we cannot imagine.

Imagine an agent analyzing market data, executing trades, managing your finances, or coordinating complex tasks with others. They will seamlessly integrate into our lives, handling everything from on-chain DeFi operations to real-world interactions, with an autonomy and intelligence far beyond what we see today.

The decentralized infrastructure (verifiable systems, payment tools, computing networks, and coordination layers) currently being built will make this future possible. For builders, investors, and enthusiasts, now is the time to dive in and shape the future.

2025 is not just a continuation, but the dawn of a new era for AI Agents.