Planet as a Brain: Towards Internet of AgentSites based on AIOS Server

The evolution of large language models (LLMs) and autonomous agents has revolutionized AI applications, enabling reasoning, memory, and tool usage. Early systems like AutoGen and MetaGPT focused on task pipelines within centralized platforms, but scalability and security issues persisted. Recent efforts such as AIOS aim to decentralize agent deployment, yet lacked robust distributed registration and discovery mechanisms. The need for a scalable, fault-tolerant infrastructure to support global autonomous agents has become urgent, especially with the rise of multi-modal AI and IoT. This research builds on prior work but introduces a novel decentralized architecture that integrates peer-to-peer protocols, addressing the limitations of existing centralized systems.

Current multi-agent systems rely heavily on centralized registries and orchestrators, which pose single points of failure, limit scalability, and raise security concerns. As the number of agents grows exponentially, maintaining real-time discovery, registration, and communication becomes increasingly challenging. Moreover, existing solutions lack resilience against network failures and malicious attacks. The core problem is designing a decentralized, scalable, secure infrastructure that supports autonomous agent registration, efficient discovery, and robust communication across global networks, ensuring high performance and safety in large-scale deployments.

The key innovations include: 1) a standardized MCP-based communication protocol supporting multi-turn, structured interactions; 2) a decentralized registration system combining DHT for scalable metadata storage and Gossip for state dissemination, avoiding reliance on centralized servers; 3) a multi-layer architecture supporting autonomous node registration, task routing, and real-time status updates; 4) integration of security features like encryption and identity validation. These innovations enable a resilient, scalable, and secure multi-agent ecosystem, facilitating autonomous collaboration at internet scale, a significant leap from prior centralized or limited-scale systems.

• �� Develop MCP v1.2.1 protocol for structured, secure communication between humans and agents, and among agents.
• �� Build a multi-layer architecture with node management, task scheduling, and monitoring modules.
• �� Implement DHT (Kademlia) for decentralized agent registration and lookup, ensuring logarithmic scalability.
• �� Use Gossip protocol for peer-to-peer state synchronization, maintaining eventual consistency.
• �� Design autonomous node registration, periodic status reporting, and task delegation workflows.
• �� Incorporate security measures including message encryption, authentication, and anomaly detection.
• �� Conduct performance evaluations through real-world deployments, measuring latency, throughput, and fault tolerance under various network conditions.

Experiments deployed the system across multiple geographic locations, simulating real-world internet conditions. Tests involved varying load levels (50, 100, 200 requests) with different concurrency (5, 10, 20 users). Metrics included response latency, throughput, registration speed, and fault recovery. Baseline comparisons with centralized architectures demonstrated superior scalability and resilience. Fault injection tests confirmed the system’s robustness, maintaining high performance despite node failures. Results validated the architecture’s capacity to support large-scale, distributed autonomous agents with low latency and high reliability.

The system maintained an average response latency of 0.15 seconds and achieved a maximum throughput of 229 requests/sec in cloud deployment. The decentralized discovery mechanism enabled rapid registration and retrieval of hundreds of agents across multiple sites, with minimal delay. Fault tolerance was confirmed through simulated node failures, where the network quickly recovered, maintaining response times and connectivity. These results demonstrate the architecture’s suitability for real-world, large-scale autonomous agent networks, outperforming traditional centralized solutions in scalability, robustness, and security.

This infrastructure supports applications like distributed AI services, smart city management, and cross-border research collaborations. It enables autonomous agents to dynamically register, discover, and communicate across multiple locations, facilitating real-time decision-making and task execution. Industries such as logistics, IoT, and finance can leverage this framework for scalable, secure, and resilient AI deployment. Long-term, it envisions a web where autonomous agents operate seamlessly across borders, forming a resilient, intelligent, distributed internet ecosystem.

Current system performance may degrade under extreme network partitioning or high latency scenarios, affecting synchronization accuracy. Security mechanisms against malicious nodes and data privacy need further enhancement. Large-scale deployment incurs significant resource costs, requiring optimization. Future work should focus on trust management, attack resistance, and efficiency improvements to ensure system robustness and wider applicability.