Networked AI Agents in a Decentralized Architecture (NANDA)¶
NANDA is building open infrastructure for the Internet of AI Agents.
As AI agents become capable of acting, communicating, and collaborating autonomously, they need infrastructure that allows them to be discovered, identified, verified, and connected across systems and organizational boundaries. NANDA develops the foundations for this networked environment.
What is NANDA?¶
NANDA is an infrastructure framework for networked AI agents.
It provides a foundation for agents to discover one another and exchange verifiable information about who they are, what they can do, and how they can be reached.
The architecture is designed to work across different agents, organizations, registries, and communication protocols rather than requiring a single system to control the entire network.
Why is this infrastructure needed?¶
The Internet was built around resources that are relatively stable and easy to identify. AI agents introduce a more dynamic environment.
Agents can change capabilities, move between deployments, operate across organizations, and interact with other agents without continuous human intervention. They therefore require infrastructure that can handle identity, discovery, capability information, trust, privacy, and dynamic communication.
NANDA focuses on these foundational requirements.
How NANDA works¶
NANDA separates stable identity and discovery from dynamic agent capabilities and contextual communication paths.
The architecture operates through a decoupled five-stage lifecycle:
| # | Stage | Component / Layer | Primary Function | Outcome |
|---|---|---|---|---|
| 1 | Identity | Persistent Identifiers | Establishes distinct, verifiable identities across organizational boundaries | Persistent Agent Handle / ID |
| 2 | Discovery | NANDA Index | Lightweight index connecting agent identifiers to distributed metadata endpoints | Discovered agent record & metadata pointers |
| 3 | Verification | AgentFacts | Structured metadata containing verifiable capabilities, endpoints, and cryptographic signatures | Cryptographically authenticated claims profile |
| 4 | Resolution | Resolution Architecture | Dynamically selects the best communication endpoint based on location, load, and context | Active, context-optimized connection endpoint |
| 5 | Interaction | Communication Protocols | Facilitates autonomous multi-agent task execution and collaboration | Interoperable cross-system execution |
A network of registries¶
Rather than relying on a single centralized directory, NANDA is designed as an open, federated network of registries.
Different organizations, ecosystems, and communities can run their own registries while remaining discoverable across the broader network.
Interoperability¶
Agents are being developed across different platforms and communication protocols. Without common infrastructure, these systems can become isolated ecosystems.
NANDA is designed to provide infrastructure across these boundaries, allowing different agent systems and protocols to participate in a broader network without requiring them to use the same internal architecture.
Verifiable identity and capability¶
Discovery alone does not establish whether an agent should be trusted.
NANDA explores mechanisms for verifiable identity, capability claims, provenance, credentials, and other evidence that can be evaluated before interaction.
This allows information about an agent to become independently verifiable rather than relying only on what an agent claims about itself.
Beyond discovery¶
The infrastructure for an Internet of AI Agents extends beyond finding and connecting agents.
| # | Focus Area | Core Research Challenge | Scope & Objectives |
|---|---|---|---|
| 1 | Interoperability | Cross-ecosystem communication | Enabling agents from different frameworks and platforms to communicate and collaborate through minimal common standards. |
| 2 | Coordination | Multi-agent alignment | Advancing from simple message exchange to synchronized action, task allocation, and decentralized coordination. |
| 3 | Evaluation | Realistic validation | Testing individual agents and agent populations under real conditions (reasoning, tool use, safety, robustness, and collective dynamics). |
| 4 | Governance | Cross-boundary oversight | Establishing verifiable identity, provenance, registry federation, oversight, and shared standards across organizational boundaries. |
What NANDA is not¶
NANDA is not an AI agent framework.
It does not define how an individual agent reasons or performs its tasks.
NANDA is not a communication protocol.
It does not replace protocols such as A2A or MCP. It provides infrastructure that can work alongside different communication systems.
NANDA is not a single centralized marketplace.
The architecture supports multiple registries and independently administered agents.
NANDA is not a universal trust authority.
It provides mechanisms for verifiable information. Different organizations and trust systems can determine how that information is evaluated and used.
Open infrastructure¶
NANDA develops open specifications, implementations, experiments, and research for the Internet of AI Agents.
The goal is to provide common infrastructure without prescribing a single way to build, operate, or govern agents.