Cryptographic X AI brings a paradigm shift: moving towards the path of the agent internet

Written by: Davide Crapis

Compiled by: Shenchao TechFlow

In recent months, the topic of "Crypto × AI (the intersection of cryptocurrency and artificial intelligence)" or "Crypto + AI" (cryptocurrency infrastructure enhanced by artificial intelligence) has garnered significant attention. Many in the blockchain community are excited about it, some are skeptical or yet to be convinced, and others are actively building it. Real-time projects at the intersection of blockchain and artificial intelligence have been improved, and many new projects are emerging.

Over the past year, I have been researching this field, particularly concerning artificial intelligence agents operating on blockchain infrastructure. Together with some colleagues from institutions like the Ethereum Foundation, Flashbots, and DeepMind, we formed a research group. We are continuing to push the boundaries of applied research to understand and test what 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 with AI agents is desirable and will produce an agent internet: an upgrade to the current paradigm of interconnectedness, enhancing incentives and modern cryptography, which will enable us to reap the benefits of an economy driven by AI agents with unprecedented security, efficiency, and collaborative potential.

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 in design and development. I will discuss their limitations and potential improvements, as well as the research needed for AI and blockchain to unlock new use cases in the medium term.

Blockchain as the Backend for the Agent Internet

Let me first say that the style of this argument will be speculative yet practical. Blockchain and artificial intelligence are two of the fastest-growing technologies of 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 evolve and interact requires some speculation. However, while the law of accelerating returns clearly points to the direction of rapid improvements, I will avoid long-term speculation on AGI. (Despite the recent hype, I believe we are relatively far from AGI that can autonomously self-improve, and it remains unclear what form they will take.)

I will focus on the short to medium-term future, where AI takes the form of human assistants and agents. In this form, AI serves as a tool for humans, facilitating the execution of human activities or performing new activities that serve humans.

Figure 1. Left: Conceptual timeline of AI evolution with increasing performance. Right: Block diagram of human activities and the activities of different forms of AI.

For years, assistants have existed in various forms, and recent advancements in LLMs indicate that the next generation of AI agents will be more capable and progress faster than before. Here is my working definition of AI agents:

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

Agents can be constrained or learn from their environment. Today, agents are often specialized for specific types of inputs and specific types of actions. For example, chatbots (like ChatGPT) take text prompts as input, may use some tools to generate answers, and respond with text output. On the other hand, trading bots take past market states as input, predict future market states and optimal actions, and execute trades. Agents can be of different types (e.g., chatbots are LLMs, while trading bots are small RL agents), and they can also be composed to perform tasks. In the future, we may find a universal architecture that can be trained to handle most use cases.

Unique and Desirable Features of Blockchain

Public blockchains have a unique set of features that make them particularly suitable for communication and interaction of AI agents. Later, we will argue that they constitute one of the best backends to support agent AI.

Decentralization: Well-designed blockchain protocols are decentralized. Furthermore, decentralization is part of the spirit of the community that originally built and upgraded them. It is built into the protocol and protected through governance.

Incentives: Well-designed blockchains have sound incentive mechanisms that drive economic security through native assets (e.g., ETH in Ethereum). Additionally, programmable smart contracts allow applications to: leverage native assets, issue new digital assets with desired properties, and define their own native assets and incentive mechanisms for their participants.

Openness and Composability: Blockchain platforms are open for user and application developer access. Moreover, applications based on smart contracts deployed on the blockchain inherit the same properties of openness and frictionless composability.

Cryptographic Guarantees: Blockchains leverage modern cryptography to provide unique levels of security, auditability, and programmable privacy. As a result, they are trust-minimized and more secure than legacy systems. Note that blockchain hacks stem from vulnerabilities in smart contracts, which are unavoidable in the early stages of the technology. As the tech stack matures, it becomes more robust and secure, while traditional systems that rely on human trust lack this attribute.

We can compare this with the traditional internet, which only has decentralization. Base protocols like TCP/IP or SMTP are open, but nearly all applications built on top of them are proprietary. This results in poor composability of the internet, which we consider a key attribute for designing agent interaction protocols. Additionally, the internet completely lacks incentives and modern cryptography at the protocol layer.

Next, we will introduce an ideal economic model where humans and agents collaborate and demonstrate how it requires the full set of features provided by blockchain protocols.

Benefits of Blockchain for AI Agents

Fast forward a few years. Assume we reach an era where AI agents can perform a large number of human activities and possess sufficient decision-making and planning capabilities. They can also autonomously execute tasks, potentially collaborating with other agents. Agents are widely deployed in society, undertaking activities that hold potentially high value for humans, whether social or financial.

Here are some properties/desires we hope these agent AI systems and their interactions with humans will possess, 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 the AI design and training processes, which largely do not directly utilize blockchain. However, the openness and composability of blockchain applications can provide unique opportunities for agent activities to be clear, automatically monitored, and attributable, which are key for incentive distribution and coordination of agent systems.

• Security: Blockchains are designed to provide reliable and secure operation in highly adversarial environments with minimal trust assumptions. Agents interacting through smart contract applications inherit these powerful attributes. Additionally, advancements in modern cryptography, such as zero-knowledge proofs, provide superpowers for smart contract applications. For example, applications can request proofs for sensitive computations while keeping agent weights and inputs private. Trustworthy smart contracts are also ideal tools for constraining agent action spaces and setting default and conditional permissions.

• Discovery: The openness of the environment between applications allows for richer request routing based on application states and the past performance of agents, all of which can be fully observed. It is easy to imagine agents credibly accumulating reputations based on their action histories, which can then be used programmatically for task ranking and discovering the best agents.

• Efficiency: Blockchain infrastructure enhances the autonomy of agents by allowing them to make important decisions, including payments, without direct human intervention, and at a low cost.

Human Desires

• Control and Programmable Privacy: Blockchain enables humans to directly own and maintain control over their agents without intermediaries. Personal data can remain private, conditionally controlling access using cryptographic tools, from fully private computation (TEE/FHE) to programmable sharing of selected attributes through zk proofs.

• Ownership and Fairness: People can establish protocols to jointly own and manage agents. Rewards for agent work can be programmatically distributed down to the smallest cent. Fairness can be measured and improved through protocol upgrades and democratic governance. The combination of blockchain infrastructure with modern identity solutions being developed can also support and automate ambitious distribution plans, such as Universal Basic Income (UBI), which is an important long-term application.

AI Supply Chain Overview

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

Global Regulation and Governance

Blockchain provides various protocols within which a wide range of rules and checks can be reliably executed. In my view, this presents a unique opportunity for global regulation of the AI market and applications, allowing for easy auditing and compliance checks. Cross-protocol transparency also makes it easy to identify deviations in real-time and deploy corrective fixes, which is impossible in traditional systems.

Risks and Costs of Blockchain Infrastructure

When training AI agents to make sensitive and impactful decisions, openness is not always desirable. For example, deploying an open-weight model for insurance underwriting decisions could expose model vulnerabilities and increase the likelihood of being attacked/exploited.

One solution might be to leverage modern cryptography to keep agents private while making their behaviors public. However, black-box adversarial machine learning attacks may still be possible, and in general, cryptographic schemes for secure yet verifiable machine learning computations are costly, adding to the already expensive training process. This is one of the most important areas of cross-research between AI security and blockchain. We need to make it technically and economically feasible in practice. A recent innovation is optimistic proofs for ML computations, which I will discuss below.

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

Blockchain-Based Systems Can Scale to Meet Demand

A common question that arises in the minds of those unfamiliar with the current state of blockchain systems is whether they are ready to accommodate the load brought by increased user activity.

For at least the past five years, this has been a focus of blockchain R&D, and today, we are at a turning point where many solutions are going live, and scalability has improved by several orders of magnitude. For example, Ethereum and its layer 2 blockchains inherit complete economic security and scalable data availability solutions, soon capable of handling tens of thousands of transactions per second (TPS). New chains are coming online that can process hundreds of thousands of transactions per second through parallelization. Shared sorting solutions and secure 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 computing and hybrid systems, making security trade-offs more effective.

As all these infrastructure innovations mature in the coming years, there is no doubt that a mature blockchain ecosystem will be able to support very high throughput, from today's tens of thousands of TPS to millions of TPS at minimal transaction costs.

The Path to the Agent Internet

The above image is a treasure map representing the three main steps on the path to the agent internet.

Let’s explore them one by one.

Enhancing Current Decentralized Applications

The first step is to enhance current blockchain applications with AI. AI has already played a role in decentralized finance (DeFi), which is by far the most popular application category. This takes the form of specialized models that continuously monitor market states to take specific actions. For example: trading bots, liquidation bots, routing bots, statistical arbitrage bots, and more broadly, bots executing strategies aimed at extracting profits from user trading flows (also known as MEV).

As the blockchain economy evolves on the current DeFi foundation, it is natural to start discussing opportunities for leveraging artificial intelligence from here.

DeFi Enhancement

Blockchain protocols are currently automated, but their interfaces are very manual, sometimes clumsy, and often inefficient. AI has the potential to become a new interface connecting humans and on-chain markets through the mediation of intelligent agents. There are at least three areas with specific opportunities to enhance current protocols.

• User Intent Matching: Users interact with AI agents to convey, and sometimes construct/refine, their intents, which AI matches to a series of on-chain actions that users delegate to it. Intent takes 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 example of intent is:

• "I want to acquire X units of token Y at no more than $Z" or

• "I want to invest $Z in Ethereum layer 2 projects every month for the next six months," or

• "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 no more than Y%."*

• While the first example requires only a few transactions, the other examples require planning, executing multiple transactions within the plan's scope, multiple price feedbacks, risk and return prediction models, and contextual information.

• Action Planning and Routing: The infrastructure for sending transactions on the Ethereum blockchain is becoming more mature and complex. There are now different routes optimized for different desires: security, speed, price efficiency, privacy. There is even a protocol aimed at making it easier to deploy new routes. Similar to what today’s DEX aggregators do for individual exchanges, more advanced routing algorithms can be designed that also consider the broader trading supply chain context and various applications. Particularly when Layer 2 applications are planning long-term strategies on behalf of users or purchasing services on Layer 1 protocols, the action space is quite large and expanding with the deployment of new mechanisms. For example, the optimal 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: Creating and managing funds where many pool resources to achieve goals and then delegate execution to AI agents. This requires aspects of intent matching and action planning, as well as the shared ownership mechanisms that blockchain can uniquely provide. For instance, a modern digital art collection agent would require all these capabilities while also leveraging the richer context provided by the latest generation of LLMs, both for integrating community preferences and for identifying assets that match them.

In all these cases, we have a primary human or community outsourcing high-value on-chain actions to some agents operating off-chain. Therefore, there is a significant demand for inference guarantees. This can be achieved in two ways:

• Running an off-chain agent network with its own security assumptions. For example, by re-mortgaging or running L1 with deliberately designed incentives, leveraging the economic security of anchor on-chain assets or ETH economic security.

• Designing an on-chain smart contract protocol for agent orchestration that requires reasoning proofs to ensure operational 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 that can economically ensure large LLM execution, which is currently impossible or prohibitively expensive using cryptographically secure zk-proofs.

AI Service Protocols

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

For example, agents serving as smart contract security auditors/testers, analysis agents, and automated financial and risk management services. Companies focused on Web3 have already provided various types of such services, but advancements in agent autonomy and reasoning proofs now offer opportunities for decentralization and eliminating 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 coordinate the agent collaboration we are now describing.

Creating New Mechanisms for Agent Services

We can leverage the superpowers of blockchain to create trusted commitment devices to implement new applications and market mechanisms that directly utilize agent users. From here, we will begin to observe the power of coordinating many agents to provide new services. We have discussed this topic in detail in a recent paper, and here I want to highlight some specific applications.

https://www.coindesk.com/consensus-magazine/2024/03/04/how-ai-crypto-will-lead-to-a-hyper-financialized-future/

AI Prediction Markets

In the short term, one of the most exciting and concrete applications is AI prediction markets. DeFi has unlocked the ability to trade long-tail assets on the blockchain, such as utility tokens of small protocols that cannot be traded in traditional markets due to the high operational costs of the infrastructure supporting them. AI prediction markets have the potential to do the same with super long-tail assets. The outcomes of the smallest events that people care about can be tokenized and traded. To make these markets work, they need:

• Effective price discovery: Including meaningful liquidity and a large volume of trades to aggregate information.

• Trustworthy market resolution: Markets need to resolve in a trustworthy and efficient manner.

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

Once these markets are operational, they become a new primitive for autonomously assessing small uncertainties without relying on a central authority, which may face security threats or biases. Various applications can be built on this foundation: micro-insurance, financial products, content moderation on decentralized social media, spam filtering, etc.

Reliable and Efficient Routing for Specialized Models

Today, most human and AI interactions are isolated in proprietary environments with general models, whether closed "cutting-edge" models (heavy models) or open-weight models (light models). However, the early success of the GPT Store and similar aggregators points to a world where these interaction patterns are just the entry point into a vast supply of GPTs, with entry points for agents with capabilities and expertise (i.e., we will soon go from explaining poker rules to playing poker, from planning trips to booking entire itineraries).

In that world, there is a clear need to efficiently route user sessions to the specialized models that can best meet their intents. When agents trade on behalf of users, there will be significant value to extract from serving users. Whether it is routers/intermediaries (extracting rents) or terminal model parties (misreporting results/performance to gain more traffic), there are incentives to extract value. Therefore, there is a clear need for trustworthy routing mechanisms and markets where service providers will compete to meet user preferences. This is an upcoming application area that I am very much looking forward to.

Creating Building Blocks for New Markets

As more specialized agents are deployed and accumulate history on-chain, building blocks for more powerful infrastructure can be developed. For example, agent discovery protocols, including reputation based on past results and agent rankings, microservices for automated bidding based on predictive results, and so on.

This is an iterative process that will take years to fully realize, with each wave of new agent service protocols creating new iterations of communication, reputation, and exchange infrastructure. The ultimate goal will be the most efficient digital coordination mechanism system, extremely bureaucratic and rent-free, which will become a pillar of the increasing share of the world economy. Ultimately, as agent capabilities continue to enhance and more real-world activities become automated, we can expect most socio-economic transactions to be resolved on this infrastructure.

Scaling Shared Ownership and Governance

Once scaled, addressing issues of shared ownership, fair value distribution, and governance of intelligent agent production systems will become crucial. Blockchain provides the foundation for achieving this solution. Today, we are in the early stages of experimentation, but some interesting models have emerged. We have two extremes:

• Direct ownership and governance minimization: This is a model that minimizes protocol governance, similar to Bitcoin. The protocol is minimal and relatively fixed. The ownership mechanisms for agent assets/resources are simple, with agent assets directly owned by their creators and value accumulated proportionally based on usage. There is a native network token that can simply be used as a utility, to pay service fees, and as a valuable capital asset for rewarding contributions.

• Shared ownership and DAO governance: The other extreme is a richer protocol, more like the applications we see today on Ethereum. There is a rich protocol specification whose parameters can be changed through explicit governance processes. 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, while the second is akin to Olas, both of which are early attempts at building an autonomous agent economy. We are still in the early stages of these new types of agent-based protocols, and new applications and capabilities may change the design of incentive and ownership models. These are just two very different examples showcasing the wide range of solutions available to protocol designers. Finally, note that similar issues exist at other levels of the AI stack beyond the agent economy, and similar solutions can be applied to incentivize AI training, data, and infrastructure services.