This repository contains six key papers outlining the architectural, strategic, and technical foundations of the NANDA (Networked Agents and Decentralized AI) initiative at MIT Media Lab. Each document contributes uniquely to the vision of the Internet of Agents.
Outlines five foundational challenges for decentralized AI systems: privacy, verifiability, incentives, orchestration, and user experience. The paper proposes a layered architectural approach and highlights parallels with internet infrastructure like TCP/IP and DNS.
Introduces the rise of intelligent agents from early LLM tooling to multi-agent systems. It discusses protocol evolution, especially the Model Context Protocol (MCP), and sets the stage for emerging agent infrastructure battles.
Explores the need for an agent registry to enable discovery, trust, and collaboration among dynamic and ephemeral AI agents. Compares governance models including platform-led, consortium-led, and decentralized blockchain-based approaches.
Details the design of a minimal, privacy-preserving registry architecture for agent discovery. Introduces the AgentFacts schema, TTL-based endpoint resolution, and cryptographic verification for agent capabilities. Offers concrete mechanisms for multi-endpoint routing, least-disclosure queries, and rapid revocation in decentralized agent ecosystems.
Analyzes the limitations of DNS and web infrastructure for AI agent systems. Weighs incremental upgrades (e.g., DNS push, SVCB records) against purpose-built registries. Offers analogies like dial-up to broadband and outlines the technical deltas introduced by the agent paradigm.
Presents a layered architecture for live multi-agent interaction via chat interfaces. Describes client, communication, and context agent roles with example use cases. Demonstrates real-time coordination enabled by the NANDA registry.
Compares leading registry architectures - MCP, A2A, Microsoft Entra Agent ID, and NANDA, across security, scalability, authentication, and maintainability. Highlights NANDA’s AgentFacts as a privacy-preserving, cryptographically verifiable schema purpose-built for dynamic, multi-agent systems.
Proposes a federated security architecture that combines NANDA’s minimal registry with the Agent Name Service (ANS) for dual-trust anchoring. Features include verifiable credentials, zero-knowledge proofs, and a modular governance system for agent discovery and capability validation.
This content is distributed under the MIT License.