Introduction
The internet has allowed seamless communication of information across individuals globally. Blockchains have enabled the coordination of individuals across the internet, as economic incentives, along with digital property rights and governance models, have fostered a new way for communities to organize and develop open source software together.
Today, as human decision making is increasingly abstracted away in favor of intelligent agents gaining rapidly more autonomy in their actions, it is becoming pivotal to ensure effective communication and trustless coordination among them.
It’s not as simple as having agents talk to each other, as we need to be able to reach consensus on shared states. Thus, these agents will need to be equipped with a set of standardized protocols and interfaces to facilitate trustless data exchange and negotiation. In order to obtain this coordination effectively, clear economic incentive structures will need to be defined to ensure strong coordination among agents to work towards common goals.
In this article, we’ll be discussing the idea of leveraging blockchain architecture to foster networks of intelligent agents. Let’s start off with a simple example — could a distributed system be built in which intelligent agents worked together to answer user-submitted questions?
In our recent project at EthGlobal Brussels, Team Zarathustra attempted to tackle this exact problem. What would the world’s first distributed inference platform even look like? We had no clue when we started, so we decided to do an experiment.
Architecture
In pursuing an alternative to a monolithic design, we decided to run a decentralized marketplace where a market of highly specialized intelligent agents could compete amongst each other to help answer questions. Models within Zarathustra perform specific computations off-chain and coordinate via a marketplace smart contract to answer user queries on-chain. By doing so, they earn economic rewards for publishing their answer in the form of an IPFS link. The IPFS links are utilized to increase efficiency, as pointing to an answer is more gas efficient than writing a 200-word answer directly to a smart contract.
You can think of this marketplace like a human brain. At any point in time, you are not utilizing 100% of your brain’s architecture. If you are reading, for example, the occipital lobe will become more active and other areas less. Similarly, we only need to activate the parts of our network that are needed to answer a user query. If a user were to ask: “what is 5+5?”, we need only to activate an arithmetic model to provide the user with an answer.
Intelligent agents in Zarathustra aren’t limited to LLMs or specific models; they can be any program with a clear definition of input and output types and their semantics. As with GPT functions, these agents can perform various tasks. This modularity fosters a diverse ecosystem of agents within an open market, each interacting through well-defined interfaces, creating a flexible and powerful computational framework, ensuring AI agents can handle tasks ranging from simple computations to complex interactions, enhancing their collective capability and adaptability.
But how does Zarathustra decide which intelligent agents need to be activated in order to solve a question? At the heart of Zarathustra’s architecture are so called ‘routers’ – a router is itself an advanced AI model whose responsibility it is to analyze queries and dispatch tasks to other appropriate specialized models for a fee. You can think of this like the master conductor of an orchestra. As the conductors, routers carefully analyze each query and then signal the appropriate specialized agents to perform their parts via event emissions.
This inter-agent communication between routers and models is one of the core concepts within Zarathustra. Each individual intelligent agent is either proficient in routing, or in specialized domains like algebra, language translation, or JavaScript programming, whereby they are activated as needed by routers. This innovative approach enables a horizontally scalable network capable of efficiently handling a diverse set of tasks.
However, it is likely that more complex questions will need more than one model to compute an accurate answer. To give an example, imagine you ask Zarathustra the question: “what is 5+5, and please translate the answer into French”. The router needs to be able to reflect different outputs between models. We’ve labeled this ‘bouncing’. Firstly, Zarathustra would utilize an arithmetic model to compute 5+5, and then ‘bounce’ the answer 10 as an input to a separate model specialized in English-French translation to retrieve the correct answer “dix”. This ‘bouncing’ enhances the system’s ability to handle diverse and complex queries in a horizontal manner.
Turing Completness
A crucial advantage of Zarathustra’s architecture is its potential for Turing completeness – the ability to perform any computable task. Zarathustra’s network of communicating agents can compute indefinitely, making dynamic decisions about when to halt processing based on their interactions and own evaluations.
This stands in contrast to current monolithic large language models, which perform fixed computations per forward pass and lack the ability to autonomously extend their processing or delve deeper into analyses. These models, while powerful, are constrained by their fixed frameworks and training data, unable to adapt dynamically to entirely new tasks without significant reconfiguration.
Zarathustra overcomes these limitations through its modular, distributed approach. The system is designed to assign computations across specialized models, scale to meet increasing computational demands, and most importantly break down complex problems into manageable sub-tasks.
The router’s ability to ‘bounce’ between models, using outputs from one as inputs for another, enables recursive, dynamic routing. This is fundamental to achieving Turing completeness, allowing Zarathustra to tackle a broad spectrum of computational challenges beyond the reach of monolithic AI systems.
Permissionless Economy
Zarathustra’s key strength lies in its permissionless nature, leveraging blockchain technology to create an open system where anyone can participate as a router or model provider. This decentralized approach, impossible with traditional payment systems, allows direct digital payments to active participants via wallet addresses. The economic incentives, driven by an open market, ensure secure, transparent, and automated transactions between users and intelligent agents. Smart contracts enable trustless coordination, allowing AI agents to collaborate without central authority, while the transparent on-chain record provides full auditability of all interactions.
Trust and accountability are thus reinforced through dual verification methods: public on-chain analysis of past inferences and user ratings. This combination significantly reduces the risk of malfeasance while fostering a competitive environment that drives innovation, where developers are incentivized to create increasingly efficient and capable models. Zarathustra thus represents a novel paradigm in AI development, combining open participation, economic incentives, and decentralized governance to create a robust, evolving ecosystem of intelligent agents.
Scalability
Unlike traditional AI systems that require extensive retraining to improve or expand their capabilities, Zarathustra’s modular architecture allows for seamless scalability and continuous improvement. New specialized models can be added to the network at any time, instantly expanding the system’s capabilities without disrupting existing functionalities or resulting in down time.
This approach enables rapid iteration and improvement. If a particular model underperforms, and its reputation decreases significantly, it can be replaced or upgraded without affecting the entire system. Similarly, as new AI breakthroughs occur in specific domains, these advancements can be quickly integrated into Zarathustra by simply adding new models or updating existing ones.
Unlike traditional monolithic AI, which relies on vertical scaling by increasing layers or parameters, horizontal scalability distributes tasks across multiple specialized agents that operate independently. As seen in other technologies, vertical scaling eventually faces diminishing returns, necessitating fragmentation and decentralization for continued growth and efficiency. We believe AI will follow this path: while vertical scaling has driven significant advancements, the future lies in a decentralized network of interoperating AI agents.
Use Cases
Zarathustra’s internal loop enhances its reasoning capabilities, allowing it to evaluate various solutions and critically analyze them. This makes it highly effective for tasks requiring exceptional thoroughness and capability, often surpassing traditional models, though it does come with increased costs and latency.
For instance, consider a complex software development project where you encounter a bug. Currently, you might use a tool like GPT-4 or Claude to identify and suggest common syntax errors based on predefined rules they have learned. However, this approach has its limitations. These tools function within a fixed computational framework, generating outputs in a single forward pass without iterative reasoning or validation. While they produce contextually appropriate responses, they lack the ability to deeply reason through complex, multi-step problems.
We propose a system that can dynamically adapt its problem-solving approach, self-evaluate, iterate over potential solutions, and apply heuristics, all within an open marketplace of intelligent agents.
For example, if you ask Zarathustra to “solve x GitHub issue,” it would utilize reasoning and iteration far beyond the capabilities of traditional models. This depth and quality might justify higher costs and longer wait times, as in many cases the correctness of the answer outweighs the speed or cost of obtaining it.
Conclusion
To us, Zarathustra represents a significant proof of concept within the field of AI, and a first step in leveraging blockchain architecture to foster networks of intelligent agents to coordinate on tasks. It’s not an approach we’ve seen taken before.
So let’s try to answer the original question when we leapt down this rabbit hole: “could a distributed system be built in which intelligent agents worked together to answer user-submitted questions?”
The answer to our experiment was a definite yes. Within 48 hours, we achieved modularity, Turing completeness, dynamic task routing, and on-chain economic incentives which served as a proof of concept of how we could in the future build out Zarathustra to handle a diverse array of computational tasks efficiently and effectively. In recognition of our hard work, we were incredibly proud and humbled to be nominated as one of the top 10 finalists for EthGlobal Brussels.
As AI continues to evolve, the integration of blockchain technology and decentralized coordination will play a crucial role in developing intelligent and adaptable systems capable of meeting the complex demands of the future, and we hope to make that a reality.
If you’d like to follow us along our journey, feel free to subscribe to our newsletter, it’s free! You can also find our Twitter here ❤️
Thomas Turek, Linkedin,
Zarathustra
