Author: Shlok Khemani, Oliver Jaros Source: Decentralised.co Translation: Shan Oppa, Golden Finance

Today’s article is an explanation of agency frameworks and our assessment of how well they develop. It is also a solicitation request, targeting founders who work in the intersection of Internet currency tracks (cryptography) and agents.

Over the past year, Decentralised.co has delved into the intersection of encryption and artificial intelligence. We even built a product that is used by over 70,000 people to track AI agents and proxy infrastructure. Although the fanaticism surrounding the field has faded in recent weeks, the impact of artificial intelligence on technology and society is something we have never seen since the Internet. If cryptocurrencies will be the financial track of the future, as we predict, its interweaving with artificial intelligence will be a recurring theme rather than a one-off.

One ​​of the more interesting categories of projects that emerge from this wave is the encryption native AI proxy framework. They are an engaging experiment that brings the core principles of blockchain – permissionless transfer of value, transparency and consistent incentives – into AI development. Their open source nature provides us with a rare opportunity to peek into their internal operations, analyzing not only their commitments, but also how they actually work.

In this article, we first analyze the actual meaning of proxy frameworks and their importance. Then, we solve an obvious question: Why do we need to encrypt native frameworks when mature options like LangChain exist? To this end, we analyze the leading crypto-native framework and its advantages and limitations in different use cases. Finally, if you are building an AI proxy, we will help you decide which framework might suit your needs. Or, should you build using a framework.

Let's go into the deeper.

Abstract

"Advanced civilization lies in expanding the number of important operations we can perform without thinking." - Alfred North Whitehead

Think about how our ancestors lived. Every family had to grow their own food, make their own clothes, and build their own shelter. They spent countless hours on basic survival tasks, and had little time to do it. Other things. Even two centuries ago, nearly 90% of people worked in the agricultural field. Today, we buy food from supermarkets, live in homes built by experts, wear clothes produced in distant factories. It once consumed several generations The task of hard work has become a simple deal. Today, only 27% of the world's population is engaged in agriculture (developed to less than 5%).

Family patterns appear when we start to master a new technology. We first understand the basics—what works, what doesn’t work, and which patterns keep appearing. Once these patterns become clear, we will They are packaged into easier, faster, and more reliable abstractions. These abstractions free up time and resources to deal with more diverse and meaningful challenges. The same is true for building software.

Taking web development as an example. In the early days, developers needed to write everything from scratch—handle HTTP requests, manage state, and create UI ——These tasks are complex and time-consuming. Then there are like React Such a framework greatly simplifies these challenges by providing useful abstractions. Mobile development follows a similar path. Initially, developers need in-depth, platform-specific knowledge until the advent of tools such as React Native and Flutter. Allow them to write code at once and deploy anywhere.

Similar abstraction patterns emerged in machine learning. In the early 2000s, researchers discovered The potential of GPUs in ML workloads. At first, developers had to fight against graphics primitives and languages ​​like OpenGL's GLSL - tools were not built for general purpose computing. In 2006, Nvidia introduced CUDA to make GPU programming more Be easier to get started and put ML Training brings to a wider community of developers, everything has changed.

As the momentum of ML development has increased, a dedicated framework has emerged to abstract GPU programming Complexity. TensorFlow and PyTorch allow developers to focus on model architectures rather than getting stuck in the quagmire of underlying GPU code or implementation details. This accelerates the iteration of model architectures and the rapidity of AI/ML we have seen in the past few years Progress.

We are now seeing similar evolution in AI agents—a software program that can make decisions and act to achieve goals, like human assistants Or the same as the employee. It uses large language models as its "brain" and can use different tools such as searching the network, making API calls, or accessing databases to complete tasks.

To build an agent from scratch, developers must write complex code to handle each aspect: how the agent thinks about the problem, how to decide what tools to use and when Use, how to interact with these tools, how to remember the context of early interactions, and how to break down large tasks into manageable steps. Each pattern must be resolved separately, resulting in repetitive work and inconsistent results.

This is where the artificial intelligence proxy framework comes in. Just as React simplifies web development by handling tricky parts of UI updates and state management, these frameworks solve common challenges in building AI agents. They provide ready-made components for the effective patterns we find, such as how to build agents’ decision-making processes, integrate different tools, and maintain contexts across multiple interactions.

Using frameworks, developers can focus on what makes their agents unique—its specific features and use cases—rather than rebuilding these basic components. Instead of months, they can create complex AI agents more easily and learn from best practices found by other developers and communities.

To better understand the importance of the framework, consider a developer who builds an agent that helps doctors review medical reports. Without a framework, they need to write all code from scratch: process email attachments, extract text from PDFs, enter text in the correct format into LLM, manage conversation history to keep track of what has been discussed, and make sure the agent responds appropriately . This is a lot of complex code for tasks that are not unique to their specific use cases.

With the proxy framework, many of these building blocks can be used directly. The framework handles reading emails and PDFs, provides patterns to build medical knowledge tips, manages dialogue flows, and even helps track important details in multiple communications. Instead of reinventing common patterns, developers can focus on aspects that make their agents unique, such as fine-tuning medical analysis tips or adding specific safety checks for diagnosis. Content that might have taken months to build from scratch can now be done in a few daysPrototyping.

LangChain has become a Swiss Army Knife developed by artificial intelligence, providing a flexible toolkit for building LLM-based applications. Although not strictly speaking, the proxy framework provides the basic building blocks for building most proxy frameworks, from chains used to sort LLM calls to memory systems used to maintain context. Its extensive integration ecosystem and extensive documentation make it the preferred starting point for developers looking to build practical AI applications.

Then are multi-agent frameworks like CrewAI and AutoGen, which enable developers to build systems where multiple AI agents work together, each agent Have its unique roles and abilities. Rather than simply performing tasks in sequence, these frameworks emphasize the collaboration of agents through dialogue to solve problems together.

For example, when assigning a research report, a proxy may outline its structure , Another agent may collect relevant information, and a third agent may comment and refine the final draft. It's like forming a virtual team where AI agents can discuss, debate and jointly improve solutions. Multi-agent systems that work together in this way to achieve high-level goals are often referred to as artificial intelligence proxy "clusters".

AutoGPT Although not a traditional framework, it pioneered the concept of autonomous artificial intelligence agents. It shows how AI accepts a high-level goal, breaks it down into subtasks, and does it independently with very little manual input. Despite its limitations, AutoGPT triggered a wave of innovation in autonomous agents and influenced the design of subsequent more structured frameworks.

But why is encryption?

All of these backgrounds eventually brought us to the rise of the crypto-native AI proxy framework. At this point, you might wonder why Web3 needs its own framework when we have relatively mature frameworks like Langchain and CrewAI in Web2? Of course, can developers use these existing frameworks to build any proxy they want? This suspicion is justified given that the industry likes to impose Web3 on any and all narratives.

We believe that there are three good reasons for the existence of Web3 specific proxy frameworks.

Financial agent running on the chain

We believe that most financial transactions will be conducted on the blockchain track in the future. This accelerates the need for a class of artificial intelligence agents that can parse on-chain data, execute blockchain transactions, and manage digital assets across multiple protocols and networks. From automated trading bots that can detect arbitrage opportunities to portfolio managers who execute earnings strategies, these agents rely on the in-depth integration of blockchain capabilities in their core workflows.

The traditional Web2 framework does not provide native components for these tasks. You have to piece together a third-party library to interact with smart contracts, parse the original on-chain events, and handle private key management - thus introducing complexity and potential vulnerabilities. Instead, a dedicated Web3 framework can handle these features out of the box, allowing developers to focus on the logic and strategies of their agents instead of fighting the low-level blockchain pipeline.

Native coordination and incentives

Blockchain involves more than just digital currencies. They provide a global, trust-minimized system of records, with built-in financial tools that enhance multi-agent coordination. Instead of relying on off-chain reputation or siloed databases, developers can use on-chain primitives such as staking, hosting, and incentive pools to coordinate the interests of multiple AI agents.

Imagine a group of agents collaborating to complete a complex task (for example, data marking for training a new model). The performance of each agent can be tracked on the chain and rewards are automatically allocated based on contributions. The transparency and immutability of blockchain-based systems allow for fair rewards, stronger reputation tracking and incentive programs for real-time development.

Encryption native frameworks can explicitly embed these features, allowing developers to design incentive structures using smart contracts without requiring proxy trust or paying to another one at a time Redesign the wheels when agents.

New Opportunities in Early Markets

Although frameworks like LangChain already have idea sharing and network effects, the field of artificial intelligence agency is still in its infancy. It is not clear what the final state of these systems will look like, and there is no way to lock in the market.

Crypto-economic incentives open up new possibilities for the way frameworks are built, managed, and monetized, which cannot be fully mapped to traditional SaaS or Web2 In economics. The experiments at this early stage can be solved for the framework itselfLock new monetization strategies, not just proxy built on top of the framework.

Contender

ElizaOS is associated with the popular project AI16Z and is a Typescript-based framework for creating, deploying, and managing AI agents. It is designed as a Web3-friendly AI proxy operating system that allows developers to build agents with unique personalities, flexible tools for blockchain interactions, and easily scale through multi-agent systems.

Rig is an open source AI proxy framework developed by Playgrounds Analytics Inc., built using the Rust programming language to create modular, scalable AI proxy. It is associated with the AI ​​Rig Complex (ARC) project.

Daydreams is a generative proxy framework that was originally created to create an autonomous proxy for on-chain games, but later extended to perform on-chain tasks.

Pippin is an AI proxy framework developed by Yohei Nakajima, founder of BabyAGI, which aims to help developers create modular and autonomous digital assistants. Yohei first built a standalone proxy and then expanded it into a common framework.

ZerePy is an open source Python framework designed to deploy autonomous agents across multiple platforms and blockchains, focusing on creative AI and social media integration. Like Pippin, Zerepy was originally an independent proxy for Zerebro, which was later expanded into a framework.

Standard

To evaluate the strength of each framework, we are from the perspective of developers who want to build AI agents. What will they care about? We think it is useful to divide evaluation into three main categories: core, feature, and developer experience.

You can think of the core of the framework as the basis for building all other agents. If the core is weak, slow, or does not evolve, the proxy created using the framework will be subject to the same limitation. Core can be evaluated based on the following criteria:

Core reasoning loop: the brain of any proxy framework; how it solves the problem. A powerful framework supports everything from basic input and output streams to complex modes such as thought chains. Without strong reasoning capabilities, the agent cannot be effectivedecompose complex tasks or evaluate multiple options, thus reducing them to gorgeous chatbots.

Memory mechanism: Agents require both short-term memory for continuous dialogue and long-term storage to acquire lasting knowledge. Good frameworks are not just remembering – they understand the relationship between different information and prioritize what information is worth keeping and what is worth forgetting.

Embedding and RAG Support: Modern agents must use external knowledge, such as documents and market data. A powerful framework can easily embed this information and retrieve it based on context through the RAG, thereby building the response on a specific knowledge rather than relying solely on basic model training.

Personal configuration: The ability to shape the communication style (tongue, etiquette and personality) of customer service personnel is crucial to user participation. A good framework can easily configure these features, recognizing that the personality of the customer service staff can significantly affect user trust.

Multi-agent coordination: A powerful framework provides built-in modes for proxy collaboration, whether through structured dialogue, task delegation, or shared memory systems. This creates a professional team, each agent exerts unique capabilities to solve problems together.

In addition to core functions, the actual utility of a framework depends to a large extent on its functionality and integration. Tools greatly expand the actual functionality of the agent. A proxy with only LLM access can participate in the conversation, but grant it access to a web browser, and it can retrieve real-time information. Connect it to your calendar API and it can schedule meetings. Each new tool adds the functionality of the proxy exponentially. From a developer's point of view, the more tools there are, the greater the optionality and scope of experimentation.

We evaluate the functionality of the encryption native framework from three dimensions:

AI model support and Features: Powerful framework provides native integration with multilingual models - from OpenAI's GPT family to open source alternatives such as Llama and Mistral. But that's not just about LLM. Support for other AI features such as text-to-speech, browser usage, image generation, and local model inference can greatly expand the capabilities of the proxy. Strong model support is becoming a must for many such frameworks.

Web3 infrastructure support: build encryptionAgents need to be deeply integrated with blockchain infrastructure. This means supporting necessary Web3 components such as wallets for transaction signatures, RPCs for chain communications, and indexers for data access. A strong framework should integrate with the fundamental tools and services across the ecosystem, from the NFT marketplace and DeFi protocols to the identity solutions and data availability layers.

Chain coverage: Web3 infrastructure support determines what agents can do, while chain coverage determines where they can do. The crypto ecosystem is developing into a decentralized multi-chain behemoth, highlighting the importance of extensive chain coverage.

Finally, even the most powerful framework can only be as good as the developer's experience. A framework can have top-notch features, but it will never be widely adopted if developers struggle to use it effectively.

The language used by the framework directly affects who can build it with it. Python has dominated both artificial intelligence and data science, so it naturally becomes the choice of AI frameworks. Frameworks written in niche languages ​​may have unique advantages, but may isolate themselves from the wider ecosystem of developers. The universality of JavaScript in web development makes it another strong contender, especially for frameworks targeting web integration.

Clear and comprehensive documentation is the lifeline for developers to adopt new frameworks. This is not just an API reference, although these are also crucial. Strong documentation includes concept overviews that explain core principles, step-by-step tutorials, well-annotated sample code, educational tutorials, troubleshooting guides, and established design patterns.

Result

The following table summarizes the performance of each framework in the parameters we just defined (ranked 1-5).

While the discussion of the reasons behind each data point is beyond the scope of this article, the following It is some outstanding impressions that each framework leaves on us.

Eliza is by far the most mature framework on this list. Given that the Eliza framework has become the Schelling point for crypto ecosystems to come into contact with AI in the recent wave of proxy, one of its prominent features is the amount of features it supports and the number of integrations.

Because of the popularity it generates, each blockchain and development tool scrambles to integrate itself into the framework (it currently has nearly 100 ) At the same time, Eliza has attracted more developer activity than most frameworks. Eliza has benefited from some very clear network effects at least in the short term. The framework is written in TypeScript, a mature language, It is further promoted by both beginners and experienced developers.

Eliza also provides it for its use of the framework for developers

We have seen a range of proxy using the Eliza framework, including Spore, Eliza (Proxy), and Pillzumi. A new version of the Eliza framework is expected to be released in the coming weeks.

Rig's approach is fundamentally different from Eliza's approach. It stands out for its powerful, lightweight and high-performance core. It supports a variety of inference modes, including prompt chains (sequential application prompts), orchestration (Coordinate multiple agents), conditional logic, and parallelism (concurrent execution of operations).

However, Rig itself is not as rich in integration. Instead, it adopts There is a different approach that the team calls “Arc handshake.” Here, the Arc team is with Different high-quality teams in Web2 and Web3 work to extend Rig’s capabilities. Some of these collaborations include working with Soulgraph to develop agent personality, and working with Listen and Solana Agent Kit to develop blockchain capabilities.

Rig has two disadvantages, though it is written in Rust, and while performing well, it is relatively few developers familiar with it. Second, we only use it in real-world applications Seeing a limited number of Rig-driven agents (AskJimmy is an exception), which makes it difficult to evaluate real developer adoption.

Before starting Daydreams , founder lordOfAFew is a major contributor to the Eliza framework. This exposed him to the growth of the framework and, more importantly, to some shortcomings. What makes Daydreams different from other frameworks isIn this way, it focuses on thinking chain reasoning to help agents achieve long-term goals. This means that when given a high-level and complex goal, the agent performs multi-step reasoning, proposes various actions, accepts or discards them based on whether they help achieve the goal, and continues the process to make progress. This makes proxy created with Daydreams truly autonomous.

The founder's background in building game projects influenced this approach. Games, especially on-chain games, are an ideal breeding ground for training agents and testing their abilities. Not surprisingly, some of the early use cases for Daydreams agents were in games like Pistols, Istarai, and PonziLand.

This framework also has powerful multi-agent collaboration and orchestration workflow implementations.

Similar to Daydreams, Pippin is also a latecomer in framework games. We cover its release in detail in this article. At the heart of Yohei’s vision is to make the agent a “digital presence” that can operate intelligently and autonomously by accessing the right tools. This vision is reflected in Pippin’s simple yet elegant core. With just a few lines of code, you can create a complex proxy that can run independently and even write code for yourself.

The disadvantage of this framework is that it even lacks support vector embedding and RAG workflows Such a basic function. It also encourages developers to use the third-party library Composio for most integrations. It is simply not mature enough compared to other frameworks discussed so far.

Some agents built with Pippin include Ditto and Telemafia.

Zerepy has a relatively simple core implementation. It effectively selects a task from a set of configured tasks and executes it if needed. However, it lacks complex inference patterns like goal-driven or thinking chain planning.

While it supports inference calls to multiple LLMs, it lacks any embedding or RAG implementation. It also lacks any primitives for memory or multi-agent coordination.

The lack of anti-reverse of this core functionality and integrationIn terms of Zerepy's adoption. We haven't seen any actual proxy using the framework go live.

Build with frameworks

If all of this sounds technical and theoretical, we I won't blame you. A simpler question is "What kind of proxy can I build with these frameworks without having to write a bunch of code myself?".

To evaluate these frameworks in practice, we identified five common proxy types that developers often want to build. They represent different levels of complexity and test various aspects of each framework's functionality.

Document Chat Agent: Tests core RAG features including document processing, context maintenance, reference accuracy, and memory management. This test reveals the framework's ability to navigate between true document understanding and simple pattern matching.

Chatbot: Assess the consistency of memory systems and behaviors. The framework must maintain consistent personality traits, remember key information in the conversation, and allow personal configurations to essentially transform stateless chatbots into persistent digital entities.

On-chain trading robot: stress-test external integration by processing real-time market data, executing cross-chain trading, analyzing social sentiment and implementing trading strategies. This reveals how the framework handles complex blockchain infrastructure and API connectivity.

Game NPC: Although the world has only started to focus on agents in the past year, agents have been playing the role of non-player characters (NPCs) in the game for decades A crucial role. Game agents are moving from rules-based agents to smart agents powered by LLM and remain the main use case for the framework. Here, we test the agent's ability to understand the environment, autonomous reasoning scenarios, and achieve long-term goals.

Voice Assistant: Evaluate real-time processing and user experience through voice processing, fast response time and messaging platform integration. This tests whether the framework can support truly interactive applications, not just simple request-response mode.

We give each framework a rating of a full 5 for each proxy type. Here are their manifestations:

Open Source Metrics

Most analytics place great emphasis on GitHub metrics when evaluating these frameworks, Such as stars and fork. Here we will give a quick look at what these metrics are and to what extent they indicate the quality of the framework.

Stars act as the most obvious signal of popularity. They are essentially bookmarks that developers give to projects they find interesting or want to track. While high star counts indicate widespread perception and interest, it can be misleading. Projects sometimes accumulate star signs through marketing rather than technical value. Think of star signs as social proofs, not quality measures.

Fork number tells you how many developers have created their own copy of the code base to build on it. More forks usually indicate that developers are actively using and extending the project. That is, many forks are eventually abandoned, so the original fork number requires context.

The number of contributors reveals how many different developers actually submitted code to the project. This usually makes more sense than star or fork. The number of regular contributors for health indicates that the project has an active community in maintaining and improving it.

We went further and designed our own indicators - contributor scores. We evaluate each developer’s public history, including their past contribution to other projects, frequency of activity and popularity of their accounts, assigning a score to each contributor. We then average all contributors to a project and weight them based on the number of contributions they make.

What do these numbers mean for our framework?

In most cases, the number of stars can be ignored. They are not meaningful indicators of adoption. The exception here is Eliza, which once became the number one trend repository of all projects on GitHub, which is consistent with being the Schelling point for all crypto AI. In addition, well-known developers like 0xCygaar have contributed to the project. This is also reflected in the number of contributors – 10 times more than other projects – Eliza attracted contributors.

Otherwise, Daydreams are very good for usFunny, simply because it attracts high-quality developers. As a latecomer launched after the peak of hype, it did not benefit from Eliza's network effects.

What's next?

If you are a developer, we hope we provide at least a starting point for you to choose which framework to build (if you need it). Beyond that, you still have to work hard to test whether the core reasoning and integration of each framework is suitable for your use case. This is inevitable.

From the observer's perspective, it is important to remember that all of these AI proxy frameworks are less than three months old. (Yes, it feels longer.) During this time, they went from being overly hyped to being called “castles in the air.” This is the essence of technology. Despite this volatility, we believe this field is an interesting and lasting new experiment in the field of crypto.

The next important thing is how these frameworks mature in technology and monetization.

In terms of technology, the biggest advantage that frameworks can create for themselves is that they enable agents to interact seamlessly on-chain. This is the primary reason why developers choose to encrypt native frameworks over general frameworks. In addition, agent and agent construction technology is a cutting-edge technical issue around the world, with new developments every day. The framework must also continue to evolve and adapt to these progress.

How the framework is monetized is more interesting. In these early days, creating a Virtuals-inspired startup platform is the at-reach fruit of the project. But we think there is a lot of room for experimentation here. We are heading towards a future with millions of agents specializing in a variety of imaginable segments. Tools that help them coordinate effectively can capture huge value from transaction fees. As a portal for builders, frameworks are certainly best suited to capture these values.

At the same time, the monetization of the framework is also disguised as the issue of monetization of open source projects and rewarding contributors who have historically engaged in free and thankless work. If a team can hack how to create a sustainable open source economy while maintaining its fundamental spirit, the impact will go far beyond the proxy framework.

These are topics we hope to explore in the coming months.