The Quilt Architecture: Decentralized Agent Discovery
How NANDA federates agent registries using push-based gossip synchronization — no central authority required.
The Centralization Trap
Every discovery system faces the same fundamental tension: centralize for simplicity, or decentralize for resilience. DNS chose a pragmatic middle ground — a hierarchical tree with root servers at the top and delegated zones below. It works, but it concentrates power in ICANN, registrars, and a handful of root server operators.
For AI agents, centralization creates unacceptable risks. A single registry controlling agent discovery becomes a chokepoint for the entire agentic ecosystem — enabling gatekeeping, censorship, rent-seeking, and single-point-of-failure outages. When Professor Raskar's team at MIT Media Lab designed the NANDA Index, they needed an architecture that was both globally coherent and fundamentally decentralized.
Their answer: the Quilt.
How the Quilt Works
The Quilt architecture treats the NANDA Index not as a single database but as a patchwork of independent registries — each operating autonomously while participating in a shared discovery fabric. Like patches in a quilt, each registry is self-contained but stitched together through standardized federation protocols.
The key design principles:
- Push-based gossip — registries push delta updates to peers rather than relying on a central coordinator, eliminating single points of failure
- Namespace autonomy — each registry controls its own namespace and registration policies
- Cross-registry resolution — any agent can be discovered from any registry through federated lookups
- Cryptographic integrity — all records are signed, preventing tampering during federation
Interactive · Quilt Federation Topology
Click any registry node · Pull-based sync — no central authority required
Registry Types in the Quilt
The Quilt accommodates diverse organizational models, each representing a different "patch" in the fabric:
| Registration Type | Example | Control Model |
|---|---|---|
| NANDA Native | @agentx | Direct registration in NANDA |
| Government | @US:shop | Location-specific sovereign domains |
| Enterprise (Routed) | @company | Access only through company registry |
| Enterprise (Direct) | @company:shop | Visible on NANDA, company-administered |
| Web3 (Routed) | @DID:company | Access through Web3 marketplace |
| Web3 (Direct) | @DID:company:agent | DID-authenticated, NANDA-visible |
This flexibility means governments can maintain sovereign control over agent registrations within their jurisdictions, enterprises can keep internal agents private while still participating in global discovery, and Web3 projects can integrate decentralized identity natively.
Why Not a Blockchain?
A natural question: why not use a blockchain for decentralized agent discovery? The NANDA team considered this and rejected it for several reasons:
- Latency — blockchain consensus takes seconds to minutes. Agent discovery needs sub-second resolution.
- Throughput — even high-performance blockchains handle thousands of transactions per second. The agentic web needs millions of lookups per second.
- Cost — gas fees or staking requirements create barriers to participation, especially for lightweight agents.
- Privacy — public blockchains expose all transactions. Agent discovery patterns are often sensitive.
The Quilt achieves decentralization through cryptographic federation rather than consensus — registries sign their records with Ed25519 keys, and peers verify signatures without needing global agreement on state. This provides the trust guarantees of decentralization without the performance penalties of blockchain.
The Lean Index Advantage
At the heart of the Quilt is the NANDA Index's lean record format. Each AgentAddr record targets approximately 120 bytes of core pointer data, containing only:
- Agent ID and human-readable name
- URLs pointing to detailed AgentFacts metadata
- Time-to-live (TTL) values
- Cryptographic signatures
All detailed metadata — capabilities, endpoints, credentials, performance metrics — lives in AgentFacts documents that can be updated independently without touching the index. This separation reduces index write overhead by approximately 10,000× compared to DNS, making global federation practical even at trillion-agent scale.
Building on the Quilt
The Quilt architecture is already operational in the NANDA testnet, with Nexartis running one of the early production nodes on Cloudflare Workers. The architecture supports:
- Cross-protocol discovery — agents registered via A2A, MCP, or HTTPS endpoints are all discoverable through the same index
- Geographic federation — registries can serve specific regions while participating in global lookups
- Enterprise isolation — companies can run private registries that selectively expose agents to the public Quilt
For a hands-on introduction to deploying agents on the NANDA network, see our NEST Quickstart tutorial.
Further Reading
- Beyond DNS: Unlocking the Internet of AI Agents — the foundational NANDA Index paper describing the Quilt federation model
- Evolution of AI Agent Registry Solutions (Aug 2025) — comparative analysis of registry architectures including federated models
- Using the NANDA Index Architecture in Practice — enterprise deployment patterns for federated registries
- Conflict-Free Replicated Data Types (Shapiro et al., 2011) — the CRDT foundations powering Quilt synchronization