Paradigm upgrade of encryption and AI integration: creating a new era of proxy Internet

Written by Davide Crapis

Compiled by: Deep Wave TechFlow

In recent months, the topic of "crypto × AI (the intersection of crypto and artificial intelligence)" or "crypto + AI" (cryptocurrency infrastructure enhanced by artificial intelligence) has been in the spotlight. Many people in the blockchain community are excited about it, some are skeptical or not convinced yet, and some are building it. Real-time projects at the intersection of blockchain and artificial intelligence have evolved, and many new projects are emerging.

Over the past year, I have been conducting research in this area, specifically regarding artificial intelligence agents running on blockchain infrastructure. Some of our colleagues at the Ethereum Foundation, Flashbots, and DeepMind, among others, have formed a research group together. We are continuing to push the boundaries of applied research to understand and test which types of AI agent applications are best suited for blockchain, and what new infrastructure we need to support them.

In this article, I will argue that integrating blockchain infrastructure and artificial intelligence agents is desirable and will result in an Internet of Agents: an upgrade to the current interconnection paradigm, with enhanced incentives and modern cryptography, that will Allowing us to reap the benefits of an economy driven by artificial intelligence agents with unprecedented security, efficiency, and potential for collaboration.

I will then discuss the path to achieving this goal. I will focus on short-term use cases and applications, some of which are already being designed and developed. I will discuss their limitations and potential improvements, as well as the research required on AI and blockchain to unlock new use cases in the medium term.

Blockchain as the backend of the proxy internet

Let me start by saying that the style of this argument will be speculative yet pragmatic. Blockchain and artificial intelligence are two of the most rapidly advancing technologies in the past decade. Both have had a profound impact on the structure of the Internet and human society. Therefore, painting a meaningful vision of how these technologies will develop and interact requires some speculation. However, while the law of expansion clearly points in the direction of rapid improvement, I would avoid long-term speculation on AGI. (Despite all the recent hype, I think we're relatively far away from autonomous self-improving AGIs, and it's unclear what form they will take.)

I will focus on the short to mid-term future where artificial intelligence takes the form of human assistants and agents. In this form, artificial intelligence is a tool that serves humanity by facilitating the performance of human activities or performing new activities that serve humanity.

Figure 1. Left: Conceptual timeline of the evolution of artificial intelligence as performance increases. Right: Block diagram of human activities and activities of different forms of artificial intelligence.

Assistants have existed in various forms over the years, and recent advances in LLMs suggest that a new generation of artificial intelligence agents will be more capable than before and progress rapidly. Here is my working definition of an AI agent:

A computer program that interacts with the world. It senses its environment through sensors (input data), processes the data autonomously (prediction and planning), and takes action to achieve goals (action).

Agents can be constrained and learn from the environment. Today, agents often specialize in specific types of input and specific types of actions. For example, a chatbot (such as ChatGPT) takes a text prompt as input, may use some tool to generate an answer, and responds with text output. Trading robots, on the other hand, take past market states as input, predict future market states and optimal actions, and execute trades. Agents can be of different types (e.g. a chatbot is an LLM and a trading bot is a small RL agent) and they can also be combined to perform tasks. In the future, we may discover a common architecture that can be trained to handle most use cases.

Blockchain has unique and desirable characteristics

Public blockchains have a unique set of characteristics that make them well suited for communication and interaction of artificial intelligence agents. We will argue later that they form one of the best backends for supporting agent AI.

• Decentralization: A well-designed blockchain protocol is decentralized. Furthermore, decentralization is part of the ethos of the communities that originally built and upgraded them. It is built into the protocol and secured through governance.
• Incentives: A well-designed blockchain has a sound incentive mechanism that drives economic security through native assets (e.g., ETH in Ethereum). In addition, programmable smart contracts allow applications to: utilize native assets, issue new digital assets with desired properties, and define their own native assets and incentive mechanisms for their participants.
• Openness and Composability: The blockchain platform is open to both users and application developers. Furthermore, applications based on smart contracts deployed on the blockchain inherit the same properties of openness and frictionless composability.
• Cryptographic Guarantees: Blockchain leverages modern cryptography to provide a unique level of security, auditability, and programmable privacy. As a result, they are trust-minimized and more secure than legacy systems. Note that blockchain hacks come from smart contract vulnerabilities, which are inevitable in the early stages of the technology. As the technology stack matures, it becomes more robust and secure, an attribute that traditional systems that rely on human trust do not have.

We can contrast this with the traditional Internet, which only has decentralization. Base layer protocols like TCP/IP or SMTP are open, but almost all applications built on top of them are proprietary. This makes the Internet less composable, which we believe is a key property for designing agent interaction protocols. Additionally, the Internet completely lacks incentives and modern cryptography at the protocol layer.

Next, we introduce an ideal economic model in which humans and agents cooperate, and show that it requires the full set of features provided by blockchain protocols.

Benefits of Blockchain for AI Agents

Fast forward a few years. Suppose we reach an era where AI agents can perform a large number of human activities and have sufficient decision-making and planning capabilities. They can also perform tasks autonomously, possibly in cooperation with other agents. Agents are widely deployed in society and undertake activities of potentially high value to humans, whether social or financial.

Here are some of the properties/desires we want these agent AI systems and their interactions with humans to have, and how blockchain can make this possible.

Agent system requirements

• Consistency: Certain aspects of agent consistency, such as value learning, explainability, and manipulability, depend on AI design and training processes that, for the most part, do not directly leverage blockchains. However, the openness and composability of blockchain applications can provide unique opportunities to make agent activity clear, automatically monitored and attributable, which is key to incentive distribution and agent system coordination.
• Security: Blockchain is designed to provide reliability and security with minimal trust assumptions in an adversarial environment where value is huge. Agents that interact through smart contract applications inherit these powerful properties. In addition, modern cryptography advances, such as zero-knowledge proofs, provide superpower for smart contract applications. For example, applications can require proof of sensitive computations, while agent weights and inputs can remain private. Trusted smart contracts are also ideal tools for limiting agents’ space for action and setting default and conditional permissions.
• Discovery: The openness of the inter-application environment allows for richer request routing based on application state and the agent's past performance, which can be fully observed. It is easy to imagine agents credibly accumulating reputations based on their history of actions, which are then used in a programmatic way to rank tasks and discover the best agents.
• Efficiency: Blockchain infrastructure enhances agent autonomy by allowing agents to perform important decisions, including payments, without direct human intervention and at low cost.

Human Desire

• Control and programmable privacy: Blockchain enables humans to directly possess and maintain control over their agents, without the need for intermediaries. Personal data can be kept private, with access conditionally controlled using cryptographic gadgets, from fully private computation (TEE/FHE) to programmable sharing of selected attributes via zk proofs.
• Ownership and fairness: People can establish agreements to jointly own and manage agents. Rewards for agent work can be assigned programmatically to as small as a penny. Fairness can be measured and improved through protocol upgrades and democratic governance. Blockchain infrastructure combined with modern identity solutions being developed can also support and automate ambitious distribution initiatives such as Universal Basic Income (or UBI), an important long-term application.

AI supply chain brief introduction

It is worth noting that in addition to communication and interoperability, blockchain infrastructure can also benefit the entire model production supply chain (data collection, data curation, training, fine-tuning). Many applications are under development, including multiple data collection protocols and computing markets. They are an important part of the decentralized AI stack, but we will not discuss them here.

Global Regulation and Governance

Blockchain provides a variety of protocols in which a wide range of rules and checks can be trusted. In my opinion, this is a unique opportunity for global regulation of AI markets and applications, allowing for easy auditing and compliance checks. Transparency across protocols also makes it easy to identify deviations and deploy corrective fixes in real time, which is not possible in legacy systems.

Risks and Costs of Blockchain Infrastructure

Openness is not always desirable when training AI agents to make sensitive and impactful decisions. For example, deploying an open weight model for insurance underwriting decisions may expose model vulnerabilities and increase the likelihood of attack/exploitation.

One solution might be to utilize modern cryptography to keep the agent private but its actions public. However, black-box adversarial machine learning attacks are still possible, and in general cryptographic schemes for secure but verifiable machine learning computations are expensive, adding overhead to the already expensive training process. This is one of the most important areas of research at the intersection of AI security and blockchain. We need to make it technically and economically feasible in practice. One recent innovation is optimistic proofs for ML computing, which I discuss below.

Another risk that has been discussed is that LLM-based oracles lower the threshold for deployment that can correctly allocate incentives to potentially harmful actions in the real world. This is not yet possible today, but more research should be done on how to enable positive use cases and how to detect and prevent harmful behavior.

Blockchain-based systems can grow to meet demand

A question that often arises in the minds of people unfamiliar with the current state of blockchain systems is whether they are ready to accommodate the load that comes with increased user activity.

This has been the focus of blockchain R&D for at least the past five years, and today we are at an inflection point with many solutions coming online and increasing scalability by orders of magnitude. For example, Ethereum and its layer 2 blockchain inherit complete economic security and scalable data availability solutions and will soon be able to handle tens of thousands of transactions per second (TPS). New chains are coming online, leveraging parallelism to process hundreds of thousands of transactions per second. Shared sequencing solutions and security bridges will allow applications deployed in different domains to interoperate securely and efficiently. Advances in zero-knowledge proof aggregation will make transactions cheaper and enable new types of off-chain computation and hybrid systems that make security tradeoffs more effective.

As all of these infrastructure innovations mature over the next few years, there is no doubt that mature blockchain ecosystems will be able to support very high throughputs, from tens of thousands of TPS per second today to extremely low per-transaction costs of hundreds. 10,000 TPS.

The Road to the Proxy Internet

The image above is a treasure map that represents the three main steps on the path to proxy internet.

Let’s explore them one by one.

Enhance current decentralized applications

The first step is to enhance current blockchain applications with AI. AI is already playing a role in decentralized finance (DeFi), which is by far the most popular application category. This takes the form of specialized models that constantly monitor the state of the market to take specific actions. For example: trading bots, liquidation bots, routing bots, statistical arbitrage bots, and more generally bots that execute strategies designed to extract profits (also known as MEV) from user trade flows.

As the blockchain economy builds on the current foundation of DeFi, it is a natural place to start discussing opportunities to leverage artificial intelligence.

DeFi enhancement

Blockchain protocols are currently automated, but interfacing with them is very manual, sometimes clunky, and often inefficient. AI has the potential to become a new interface connecting humans and on-chain markets, mediated by intelligent agents. There are at least three areas where there are specific opportunities to enhance current protocols.

• User intent matching: Users interact with AI agents to communicate and sometimes construct/refine their intentions, and AI matches them to a series of on-chain actions entrusted to it by the user. Intentions take the form of a goal and multiple safeguards, and actions can be a single transaction or a structured plan executed over a longer time scale. A simple intent example is

1. “I want to get X units of token Y at a price no higher than $Z” or
2. “I want to invest $Z per month in Ethereum Layer 2 projects for the next six months”, or
3. “I want to re-stake my $ETH to EigenLayer and delegate it to AVSs with an APR of at least X% and a risk factor of at most Y%”.

• While the first example only requires a few trades, other examples require a plan, a plan to execute multiple trades within the plan, multiple price feeds, predictive models of risk and reward, and contextual information.
• Action planning and routing: The infrastructure for sending transactions on the Ethereum blockchain has become more mature and sophisticated. Now there are different routes optimized for different desires: security, speed, price efficiency, privacy. There's even a protocol designed to make deploying new routes easier. Similar to what today's DEX aggregators do for individual exchanges, more advanced routing algorithms can be designed that also take into account the broader trading supply chain context and various applications. Especially when planning long-term strategies on behalf of users or Layer 2 applications purchasing services on Layer 1 protocols, the room for action is considerable and is expanding as new mechanisms are deployed. For example, the best plan for user portfolio optimization might be to partially redeploy their funds to a cheaper Layer 2 and execute their investments there.
• Shared funds and asset pools: Create and manage funds where many people pool resources, achieve goals, and then delegate execution to AI agents. This requires aspects of intent matching and action planning, as well as shared ownership mechanisms that blockchain can uniquely provide. For example, a modern version of a digital art collection agent will need all of these capabilities and also take advantage of the richer context provided by the latest generation of LLMs, both for synthesizing community preferences and identifying assets that match them.

In all of these cases, we have a dominant human or community outsourcing high-value on-chain actions to some agent running off-chain. Therefore, there is a large need for inference guarantees. This can be achieved in two ways:

• Run a proxy network off-chain, with its own security assumptions. For example, take advantage of the economic security of assets on the anchor chain or the economic security of ETH by re-staking or running L1 with specially designed incentives.
• Use on-chain smart contracts to design agent orchestration protocols that require inference proofs to ensure operation validity. This can be achieved through zkML (zk-proofs) or opML (optimistic proofs). Both areas are progressing rapidly, but opML is a very interesting solution to economically secure large-scale LLM executions, which is impossible or cost-prohibitive today using cryptographically secure zk-proofs.

AI Service Agreement

A related category is enhancing protocol infrastructure with autonomous agents rather than retail applications. Most of the applications here are similar to agent-based products built for traditional business services, but these agents can take advantage of the openness, liveness, and data richness of blockchain.

For example, agents acting as smart contract security auditors/testers, analytics agents, and automated financial and risk management services. Web3-focused companies already provide various types of such services, but advances in agent autonomy and proof-of-inference now offer the opportunity to decentralize and remove trust from critical services to protocol operations.

A new application area is content management. With the rise of decentralized social media like Farcaster and Lens, new opportunities for agent automation/intermediary management have emerged. However, these require the creation of new mechanisms to orchestrate the agent collaboration we now describe.

Create a new agency service mechanism

We can leverage blockchain’s superpower to create trusted commitment devices to implement new applications and new market mechanisms that directly leverage agent users. From here we will start looking at the power of coordinating many agents to provide new services. We discussed this topic in detail in our recent paper, and here I want to focus on some specific applications.

AI prediction market

The most exciting and concrete application in the short term is AI prediction markets. DeFi unlocks the ability to trade long-tail assets on the blockchain, such as utility tokens of small protocols, which cannot be traded in traditional markets because the infrastructure to support them is too expensive to operate. AI prediction markets have the potential to do the same thing with ultra-long-tail assets. The results of the smallest events that people care about can be tokenized and traded. For these markets to work, they need:

• Efficient price discovery: Includes meaningful liquidity and large volumes to aggregate information.
• Credible market solutions: Markets require credible and efficient solutions.

AI can automate these operations by having professional trading agents query LLMs to obtain probability estimates of events and then place bets, as seen in recent large-scale competitions. It has also been suggested that multi-round dispute protocols could be used for automated market resolution, using LLM in early rounds and only involving humans in cases that escalate to later rounds.

Once these markets work, they become a new primitive for evaluating small uncertainties with complete autonomy, without relying on a central authority that may face security threats or bias. Various applications can be built on this basis: microinsurance, financial products, content moderation on decentralized social media, spam filtering, etc.

Provide reliable and efficient routing for specialized models

Today, most human and AI interactions are isolated in proprietary environments with common models, whether closed “frontier” models (heavy models) or open weight models (light models). However, the early success of the GPT Store, and aggregators alike, points to a world where the above interactive model is just the entry point into a vast GPT supply with agency capabilities and expertise (i.e., we will soon go from explaining the rules of poker to Play poker, from planning to booking the entire trip).

In that world, there is a clear need to efficiently route user sessions to the specialized model that best serves their intent. When agents conduct transactions on behalf of users, there is a significant amount of value that can be extracted from service users. Whether it's the router/intermediary side (extracting rent) or the endpoint model side (false positive results/performance to get more traffic), there is an incentive to extract value. Therefore, there is a clear need for a trusted routing mechanism and a market where service providers will compete to meet user preferences. This is an upcoming application area that I am very excited about.

Create building blocks for new markets

As more agents with specialized skills are deployed and accumulate history on the chain, the building blocks of a more robust infrastructure can be developed. For example, agent discovery protocols, including reputation based on past results and agent rankings, automated bidding of microservices based on predicted results, and more.

This is an iterative process that will take years to fully implement, with new iterations of communications, reputation, and exchange infrastructure evolving as each new wave of proxy service protocols is created. The ultimate goal will be the most efficient system of digital coordination mechanisms, extremely convenient and rent-free, which will become the backbone of an ever-increasing share of the world economy. Ultimately, as agent capabilities continue to increase and more real-world activities are automated, we can expect that the majority of socioeconomic transactions will be resolved on this infrastructure.

Extending Shared Ownership and Governance

Once at scale, addressing issues such as shared ownership, fair value distribution, and governance of smart agent production systems will become critical. Blockchain provides the basis for implementing this solution. Today we are in the early stages of experimentation, but some interesting models are emerging. We have two extremes:

• Direct ownership and minimal governance: This is a model where protocol governance is minimized, similar to Bitcoin. The protocol is minimal and relatively fixed. The proxy asset/resource ownership mechanism is simple, proxy assets are owned directly by their creators and accumulate value in proportion to their usage. There is a native network token that can be used simply as a utility, payment for services, and as a valuable capital asset to reward contributions
• Shared ownership and DAO governance: The other extreme is a richer protocol more like what we see on Ethereum today. There is a rich protocol specification whose parameters can be changed through an explicit governance process. Native tokens can be used for governance and have richer incentive mechanisms that enable shared ownership of different system components.

The first is similar to what Morpheus is experimenting with, and the second is similar to Olas, both early attempts at building an autonomous agent economy. We are still in the early stages of these new types of agent-based protocols, and there will be new applications and new capabilities that may change how incentive and ownership models are designed. These are just two very different examples that illustrate the wide range of solutions available to protocol designers. Finally, note that similar problems exist at other levels of the AI stack beyond the agent economy, and similar solutions can be used to incentivize AI training, data, and infrastructure services.