Planet as a Brain: Towards Internet of AgentSites based on AIOS Server (2504.14411v3)

Abstract: The internet is undergoing a historical transformation from the "Internet of Websites" to the "Internet of AgentSites." While traditional Websites served as the foundation for information hosting and dissemination, a new frontier is emerging where AgentSites serve as the hubs of the internet, where each AgentSite hosts one or more AI agents that receive tasks, address them, and deliver actionable solutions, marking a significant shift in the digital landscape and representing the next generation of online ecosystems. Under this vision, AIOS, the AI Agent Operating System, serves as the server for the development, deployment and execution of AI agents, which is a fundamental infrastructure for the Internet of Agentsites. In this paper, we introduce AIOS Server, a runtime framework to host agents and enable global-scale collaboration among decentralized agents. AIOS Server provides a communication protocol leveraging the Model Context Protocol (MCP) and JSON-RPC to enable agent-agent or human-agent interactions. Each AIOS node operates as a server to host and execute agents, while supporting peer-to-peer coordination without reliance on centralized orchestration. Based on AIOS Server, we further present the world's first practically deployed Internet of Agentsites (AIOS-IoA), including AgentHub for agent registration and discovery and AgentChat for interactive communication, at https://planet.aios.foundation. The agent discovery mechanism based on Distributed Hash Tables (DHT) and a Gossip protocol serves as the search engine for the internet of agentsites. This work provides a practical foundation for building the Internet of Agentsites-a new paradigm where autonomous agents become first-class citizens of the web. The implementation is available at https://github.com/agiresearch/AIOS.Server and is integrated into the AIOS main branch at https://github.com/agiresearch/AIOS.

• The paper presents a novel decentralized framework that shifts from static websites to dynamic AgentSites hosting autonomous AI agents.
• It establishes standardized JSON-RPC protocols for both human-agent and agent-agent communications to streamline task delegation and collaboration.
• Experimental evaluations reveal low latency and efficient node discovery using Kademlia DHT and Gossip protocols, demonstrating promising scalability.

This paper introduces the concept of the "Internet of AgentSites," proposing a shift from the current Internet of Websites, where static information is hosted, to a future where decentralized AgentSites, each running AI agents, serve as hubs for task execution and collaborative problem-solving. The core infrastructure enabling this vision is the AIOS (AI Agent Operating System) Server, presented as a runtime environment for developing, deploying, and executing AI agents in a distributed manner.

The paper addresses the limitations of existing agent systems, which are often confined to centralized platforms, hindering their openness, interoperability, and scalability in multi-agent ecosystems. AIOS Server is designed as a decentralized framework where each node operates independently, capable of hosting agents and facilitating peer-to-peer coordination without centralized orchestration.

A key component of AIOS Server is its standardized communication protocol, built upon the Model Context Protocol (MCP) and JSON-RPC. This protocol governs interactions between humans and agents (Human-Agent Communication) and between agents themselves (Agent-Agent Communication).

• Human-Agent Communication: This protocol allows human users to interact with agents by sending structured requests (tasks or queries) and receiving structured responses. The workflow involves task initialization via a JSON-RPC request, agent processing, response generation, and optional iterative refinement. The message format is standardized using JSON-RPC fields like jsonrpc, id, method, and params, which include sender, recipient, messages, and constraints like maxTokens. The agent's response mirrors this structure, including sender, recipient, content, model used, and stopReason. An example request is provided where a human user asks an "academic_agent" to summarize a paper, and the agent responds with the summary.
• Agent-Agent Communication: This protocol enables agents to interact with each other for task delegation, data sharing, and collaborative workflow execution in a distributed environment. The workflow involves agent discovery, task delegation (where one agent sends a task request to another), task execution by the receiving agent, and response handling by the originating agent. Messages are structured JSON-RPC requests containing intent, sender, recipient, and a task object with name and arguments. Responses include the task result with status, output, and an isError flag. An example shows an "analysis_agent" delegating a data extraction task to a "data_agent."

The paper highlights the similarities (structured messaging via JSON-RPC, intent specification) and differences (initiator, message flow, interaction type, routing) between these two protocols.

To support scalable agent communication in a decentralized network, AIOS Server implements a robust node registration and discovery mechanism. Besides nodes that host agents, the system includes one or more agent registry nodes. When an AIOS Server node comes online, it registers itself and its hosted agents with selected registry nodes. Nodes periodically broadcast their availability and capabilities using structured metadata (e.g., agent_id, description, last_seen).

The decentralized discovery mechanism leverages a combination of a Kademlia-based Distributed Hash Table (DHT) for structured storage and lookup ($O(\log n)$ complexity) and a Gossip protocol for periodic peer-to-peer propagation of agent presence and status updates (eventual consistency). This hybrid approach ensures fault tolerance and resilience against node churn. The process involves agent launch and local registration, registration on DHT for replication, Gossip dissemination of status deltas, and state synchronization among nodes. The paper provides Python pseudocode excerpts demonstrating the core logic for the DHT class, Gossip protocol, Agent Directory Service, and the Gossip Integrator.

Each AIOS Server node acts autonomously, capable of hosting local agents and delegating tasks to agents on other nodes. When a task arrives, the node first attempts local execution. If no suitable local agent is available, it delegates the task to another node discovered through the decentralized registry, enabling dynamic load balancing and fault tolerance. Nodes periodically report their status, including system information (CPU/memory usage, platform) and available agents, using a defined JSON format.

The paper presents the "world's first practically deployed Internet of Agentsites (AIOS-IoA)" at \url{https://planet.aios.foundation}. This platform includes components like AgentHub for agent registration and management and AgentChat for human-agent interaction. The paper includes visualizations of the deployed network, showing nodes in different geographic locations and user interface examples demonstrating the node dashboard, single-node details, and task execution logs, validating the practical implementation.

Experimental evaluation focused on communication performance (latency and throughput) and node discovery efficiency. Tests under local and cloud environments with varying loads demonstrated low latency (under 200ms) and predictable throughput scaling. The decentralized node discovery mechanism showed rapid registration, with average latency remaining at 1ms across different numbers of nodes (3, 5, 7). These results support the system's viability for decentralized agent communication and management.

The authors conclude that AIOS server provides a practical foundation for the Internet of AgentSites, highlighting its modular architecture, structured communication, and efficient decentralized mechanisms. Future work includes optimizing inter-node communication, developing adaptive load balancing, enhancing resilience under partial failures, validating the DHT registry at larger scales, and designing security mechanisms for peer-to-peer agent communication.

The implementation is available at \url{https://github.com/agiresearch/AIOS.Server} and will be integrated into the main AIOS repository at \url{https://github.com/agiresearch/AIOS}.