Discover the Future: A Guide to the Internet of Agents

In recent months, the fusion of cryptocurrency and artificial intelligence (AI) has become a hot topic within the blockchain community. While some enthusiasts are optimistic and actively developing projects at this intersection, others remain skeptical or unconvinced. Despite varying opinions, the momentum behind AI agents operating on blockchain infrastructure continues to grow, as evidenced by revamped existing projects and new initiatives.

We explored the practical applications and potential of AI agents within blockchain ecosystems. We aim to understand and test which AI agent applications are best suited for blockchains and what new infrastructure is necessary to support them.

In this article, we argue that integrating blockchain infrastructure with AI agents is desirable and will lead to the emergence of an "Internet of Agents." This new paradigm promises enhanced security, efficiency, and collaboration, driven by AI agents operating within an economy underpinned by modern cryptography and incentives. We will outline the steps needed to achieve this vision, focusing on short-term use cases and applications currently in development. We will also discuss their limitations, potential improvements, and research required to unlock new medium-term use cases.

How Does an AI Agent Work?

An AI agent operates by defining, creating, or inferring an objective and breaking it down into tasks to complete. This process can be divided into three stages: task definition and planning, decision-making, and feedback and adaptation.

Task Definition and Planning

Define and Assign an Objective: Provide the agent with a specific goal to accomplish.
Assign Resources: Select the tools and information sources the agent will use to achieve its goal.
Environmental Assessment: The agent collects information about its environment using sensors and other data sources to understand its task's context and potential obstacles.
Plan Generation: Considering the tools available, the agent devises strategies to achieve its goal by breaking it down into subtasks and subgoals.

Decision-Making

Data Analysis: The agent analyzes available data, such as environmental sensor readings, past experiences, and its model, to predict the outcomes of potential actions.
Action Execution: The agent selects and executes the actions that maximize the likelihood of success.

Feedback and Adaptation

Performance Monitoring: The agent monitors the outcomes of its actions to evaluate their effectiveness in accomplishing the objective.
Feedback Loop: The agent uses feedback to adjust its strategy and actions. If programmed to do so, it can request human intervention if it encounters difficulties.
Adaptation and Learning: The agent continuously learns from its experiences, updating its knowledge base and decision-making processes based on new information.

Components of an AI Agent

AI Model: The core decision-making mechanism, often using recent advancements like LLMs (large language models), VLMs (vision-language models), or LMMs (large multi-modal models). The model processes data, makes decisions, and takes actions to achieve the agent’s goals.
Sensors: These components collect data from the agent’s environment, enabling it to "perceive" accurately and act accordingly. In software agents, sensors might be digital interfaces to websites or databases, while in robotic agents, they include cameras and microphones.
Actuators: How the agent takes actions based on its objectives and data collected from sensors. In software agents, actuators control other applications or devices, while in robotic agents, they could be limbs, display screens, or speakers.

Multi-Agent Systems

A Multi-Agent System (MAS) allows multiple AI agents to collaborate to solve problems and achieve goals, offering several advantages over single-agent systems:

Increased Complexity Handling: Agents can learn from the behavior of other agents and their environment. MAS is scalable, allowing additional agents to be added if the existing system can't handle the task or growing demand.
Fault Tolerance: MAS is more resilient to individual agent failures, providing higher availability, which is critical for systems representing essential functions.
Cooperation: Agents in a MAS can collaborate to achieve goals, sharing information about their actions to avoid redundant efforts.

Blockchains as the Backend for the Internet of Agents

Our argument is both speculative and practical. Blockchain and AI have been the most rapidly advancing technologies over the past decade, significantly impacting the internet and society. While speculation is necessary to envision their future interactions, we will avoid long-term predictions about autonomous general intelligence (AGI). Instead, we will focus on the immediate future, where AI functions as human assistants and agents.

AI agents, defined as computer programs that interact with the world by perceiving their environment, processing data autonomously, and taking actions to achieve goals, are becoming increasingly capable. While today's agents are often specialized (e.g., chatbots like ChatGPT or trading bots), future agents may adopt more general architectures capable of handling diverse use cases.

Unique Features of Blockchains

Public blockchains possess a unique combination of features that make them ideal for supporting AI agents:

Decentralization: Blockchain protocols are inherently decentralized, with governance and upgrades built into their design.
Incentives: Blockchains have built-in incentive mechanisms that drive economic security through native assets and programmable smart contracts.
Openness and Composability: Blockchains provide open access for users and developers, enabling frictionless composability of applications.
Cryptographic Guarantees: Modern cryptography ensures high levels of security, auditability, and privacy, making blockchains more trust-minimized than traditional systems.

Unlike the legacy internet, which lacks incentives and modern cryptography at the protocol level, blockchains offer a comprehensive suite of features essential for designing protocols for agent interaction.

Benefits of Blockchains for AI Agents

Envisioning a future where AI agents autonomously execute a wide array of human activities and collaborate with other agents, blockchains offer several benefits:

Alignment: Blockchains can enhance agent alignment through transparency, automated monitoring, and incentive distribution.
Safety: Blockchains are designed for reliability and security in adversarial environments. Advances in cryptography, such as zero-knowledge proofs, can further enhance agent safety by enabling private but verifiable computations.
Discovery: Open environments on blockchains allow for richer routing of requests and credible reputation accumulation based on agents' action history.
Efficiency: Blockchain infrastructure supports autonomous agent decisions, including payments, with minimal human intervention and low costs.

For humans, blockchains provide control, privacy, ownership, and fairness through direct agent management, collective ownership protocols, and programmable incentive mechanisms.

AI Supply Chain and Global Regulation

Beyond agent communication and interoperability, blockchains can benefit the AI supply chain, including data gathering, curation, training, and fine-tuning. Additionally, blockchains offer versatile protocols for enforcing global AI market regulations, providing transparency, and enabling real-time corrective actions.

Risks and Costs

While blockchains offer numerous benefits, they also present risks. Openness can expose model vulnerabilities, and cryptographic schemes for secure computations are costly. Moreover, deploying AI agents for sensitive decisions (e.g., insurance underwriting) may lead to adversarial attacks. Addressing these challenges through ongoing research is crucial.

Scalability of Blockchain Infrastructure

Concerns about blockchain scalability are being addressed through various solutions, such as Layer 2 blockchains and parallel processing chains. These advancements will enable blockchain ecosystems to support high throughput and low-cost transactions, making them ready for widespread AI agent deployment.

Path to the Internet of Agents

Achieving the Internet of Agents involves three main steps:

Augmenting Current Decentralized Applications: AI can enhance decentralized finance (DeFi) by improving user intent matching, action planning, and shared funds management.
AI Services to Protocols: AI agents can provide autonomous services like security auditing, analytics, and treasury management, leveraging blockchain transparency and data abundance.
Creating New Mechanisms for Agent Services: Blockchain commitment devices can implement new applications and market mechanisms, such as AI prediction markets and efficient routing to specialized models.

As more agents with specialized skills are deployed and accumulate on-chain history, new market infrastructures for communication, reputation, and exchange will evolve, forming the backbone of a significant portion of the global economy.

Shared Ownership and Governance

Solving issues of shared ownership, equitable value distribution, and governance of agent systems is critical. Blockchain protocols can enable diverse solutions, from minimal governance to rich protocol specifications with democratic processes. These models will evolve with AI agents' increasing capabilities and applications.

In conclusion, integrating blockchain infrastructure and AI agents can revolutionize digital coordination and economic exchange, paving the way for an Internet of Agents with unprecedented security, efficiency, and collaborative potential.