Usable Agent Discovery

• Usable Agent Discovery is a framework that enables autonomous agents to be efficiently located, identified, and contextually matched through cryptographically verifiable credentials and dynamic endpoint resolution.
• It leverages decoupled identity models, hierarchical naming schemes, and layered registries to ensure migration-invariant references and secure metadata representation.
• Discovery protocols combine semantic search, adaptive ranking, and privacy-preserving techniques to achieve low-latency, scalable, and resilient agent resolution in distributed environments.

Usable Agent Discovery refers to the set of architectural, protocol, and data modeling principles and mechanisms that enable autonomous software agents to be located, identified, and contextually matched in large-scale, heterogeneous multi-agent systems. Effective agent discovery encompasses not only the retrieval of candidate agents by capabilities or roles but also cryptographically assured identity, dynamic endpoint resolution, context-aware ranking, and scalable, interoperable workflows. This foundation is critical for distributed AI ecosystems, Internet of Agents (IoA) infrastructures, and AI-enabled service platforms.

1. Architectural Patterns and Naming Schemes

Modern approaches to usable agent discovery universally separate the concerns of identity, discovery, and dynamic endpoint resolution. Key patterns include:

• Decoupling Identity from Location: The agent:// URI scheme introduces a topology-independent identifier composed of a trust-root domain, hierarchical capability path, and sortable unique ID. This design ensures migration-invariant references, removing the brittle dependence of agent identity on network endpoints (Rodriguez, 21 Jan 2026).
• Hierarchical, Capability-Indexed Naming: Systems such as AgentDNS and the NANDA Index utilize structured, multi-tiered namespaces (e.g., agentdns://org/cat/name or URN-style agent names) to encode semantic, provider, and organizational boundaries and support hierarchical or prefix search (Cui et al., 28 May 2025, Raskar et al., 18 Jul 2025). Names typically map to registries that expose statically verifiable metadata and capability manifests.
• Layered Registry Architecture: Architectures like NANDA and ADS interpose lean, federated registries, which map agent names or capability keys to metadata endpoints, dynamic resolvers, and cryptographically verifiable credentials (Raskar et al., 18 Jul 2025, Muscariello et al., 23 Sep 2025).

2. Metadata Representation, Credentialing, and Capability Modeling

Agent discoverability demands richly structured metadata:

• AgentFacts and Verifiable Credentials: Typical agent metadata cards encapsulate URN identity, functional capabilities, precise resource and context requirements (e.g., geographic region, trust level, latency SLO), cryptographic keys, endpoints, and attestation objects (Ed25519/PASETO/W3C VC) (Zinky et al., 5 Aug 2025, Raskar et al., 18 Jul 2025, Rodriguez, 21 Jan 2026).
• Schema Validation and Extensibility: Frameworks such as Open Agentic Schema Framework (OASF) and NANDA’s JSON-LD-based AgentFacts enforce both schema-level normalization (capabilities, endpoints, resource claims, etc.) and upgradable extensibility for emerging agent types (e.g., LLM prompt agents, ACP servers) (Muscariello et al., 23 Sep 2025, Raskar et al., 18 Jul 2025).
• Cryptographically Verifiable Capability Claims: Agent entries are bound to cryptographically signed claims (VCs, PASETO tokens, Sigstore signatures). Attestation covers identity, asserted skill path, organizational trust root, and may embed zero-knowledge proofs for privacy-preserving validation (Rodriguez, 21 Jan 2026, Raskar et al., 18 Jul 2025, Mittal et al., 29 Apr 2026).

3. Discovery Mechanisms, Resolution Protocols, and Ranking

Usable agent discovery workloads must support (1) contextually relevant retrieval, (2) scalable, rapid lookup, (3) dynamic endpoint selection, and (4) resilience to churn.

• Registry/Discovery Service: Core CRUD operations over agent registries are exposed via RESTful or gRPC APIs. Implementation patterns range from centralized root registries (AgentDNS, ANS), federated CRDT-backed indexes (NANDA), to DHT-based multi-level mappings (AGNTCY ADS) (Cui et al., 28 May 2025, Raskar et al., 18 Jul 2025, Muscariello et al., 23 Sep 2025, Mittal et al., 29 Apr 2026).
• Semantic and Scored Search: Retrieval functions combine syntactic (inverted index over tags/skills), semantic (dense vector embedding or hybrid keyword/embedding RAG), and policy-aware matching (e.g., scoring on capability, resource, and trust) (Xu, 4 May 2026, Guo et al., 24 Nov 2025, Zinky et al., 5 Aug 2025).
• Multi-Criteria Contextual Ranking: Adaptive discovery protocols employ weighted scoring over contextual dimensions (e.g., proximity, agent/system load, trust level) or composite ranking incorporating both semantic match and live state metrics. Example:

$$\mathrm{Score}_i = w_{geo}\,e^{-d_{i,\mathrm{geo}}/D_{\max}} + w_{load}\left(1 - \frac{L_i}{L_{\max}}\right) + w_{trust}\mathrm{TrustLevel}_i, \;\; w_{geo} + w_{load} + w_{trust} = 1$$

(Zinky et al., 5 Aug 2025)

• Dynamic Resolution and Adaptive Endpoints: Rather than returning static URLs, resolvers return context-tailored endpoints based on the requester's live context, negotiation outcome, and global system state. Real-time context updates trigger re-resolution and session migration (Zinky et al., 5 Aug 2025, Raskar et al., 18 Jul 2025).

4. Security, Privacy, and Trust Models

Discovery must provide both integrity guarantees and enforced access boundaries:

• Decentralized Identifiers (DIDs) and PKI Integration: Protocols such as ANS use W3C DIDs, agent-specific certificates, and signature chains issued by CA roots for authentication (Mittal et al., 29 Apr 2026).
• Verifiable Credentials and Zero-Knowledge Proofs: Capability VCs, often accompanied by ZK-proofs (e.g., SNARKs) or least-disclosure VC flows, enable cryptographic attestation without exposing sensitive secrets or privileged policies (Raskar et al., 18 Jul 2025, Mittal et al., 29 Apr 2026).
• Admission and Policy Controls: Kubernetes-native ANS leverages CRDs, admission webhooks, and policy-as-code (e.g., Open Policy Agent) to validate naming, capability bounds, and resource constraints at deployment time (Mittal et al., 29 Apr 2026).
• Privacy-Preserving Queries: Mechanisms such as privacy-path queries, mix-nets, and minimal disclosure protocols are used to restrict metadata leakage and support cross-organizational privacy (Raskar et al., 18 Jul 2025, Cui et al., 28 May 2025).

5. Scalability, Resilience, and Performance Engineering

Demands of web-scale, heterogeneous AI agent ecosystems have driven several performance-focused innovations:

• Indexing and Lookup Complexity: Systems leverage Kademlia DHTs for $O(\log N)$ lookup (agent:// schemes, ADS, NANDA), CRDTs for index convergence with $<100$ ms lag, and product-quantization for dense capability indexing (Rodriguez, 21 Jan 2026, Guo et al., 24 Nov 2025, Raskar et al., 18 Jul 2025).
• Microservice and Caching Architecture: Stateless resolvers, TTL-conditioned caches, and horizontally sharded authoritative servers provide $O(1)$ mean network call performance in steady state (Zinky et al., 5 Aug 2025).
• Latency and Throughput Benchmarks: Modern registry and discovery systems typically achieve agent fact lookups and context-aware resolution in $<30$–$60$ ms (NANDA, AdaptiveResolver), agent registration and discovery in $<45$ ms and $<12$ ms (ANS on Kubernetes), and throughput up to $10^4$–$5\times10^4$ resolution requests/sec per cluster (Zinky et al., 5 Aug 2025, Mittal et al., 29 Apr 2026, Raskar et al., 18 Jul 2025).
• Churn and Overlay Design: Peer-to-peer overlays (structured DHT vs. gossip protocols) are benchmarked under both node-level and agent-level churn (e.g., warm/cold life cycles), with structured overlays (Kademlia) offering superior resilience in node-failure regimes and gossip-based overlays excelling when agent warm-up delays dominate (Dazzi et al., 25 Apr 2026).

6. Practical Workflows and Implementations

Usable agent discovery encompasses both infrastructural protocols and practical developer-facing workflows:

• End-to-End Query/Invocation Patterns: Agents typically discover registry/endpoint records, resolve for dynamic connection parameters (QoS, placement, authentication), and may negotiate communication specs via multi-phase protocols before invoking application-level interactions (Zinky et al., 5 Aug 2025, Cui et al., 28 May 2025).
• SDK and API Integration: Python and TypeScript/Go SDKs encapsulate registration, discovery, signature verification, and policy handling (e.g., AGNTCY ADS, ANS), abstracting protocol complexity for application developers (Muscariello et al., 23 Sep 2025, Mittal et al., 29 Apr 2026).
• Autonomous Capability Announcement: Agents periodically publish structured, machine-interpretable capability profiles (skills, tools, constraints, signed claims), supporting continuous index updates and enabling autonomous discovery and composition (Guo et al., 24 Nov 2025).

7. Evaluation Results, Challenges, and Open Directions

Empirical evaluations and reported metrics demonstrate high effectiveness but identify persistent open problems:

• Precision and Recall: Systems such as agent:// URIs demonstrate $O(\log N)$0 discovery, while GRAIL achieves Recall@10 of $O(\log N)$1—a large margin over monolithic dense retrieval (Rodriguez, 21 Jan 2026, Xu, 4 May 2026).
• Resilience and Usable Availability: Structured overlays yield near-perfect usable availability under moderate churn, but agent/host life cycles introduce complex trade-offs favoring different overlays depending on workloads and SLOs (Dazzi et al., 25 Apr 2026).
• Limitations and Research Frontiers: Current limitations include centralization bottlenecks (single roots), evolving schema/ontology alignment, limited coverage of multimodal capabilities, and incentive-compatible federated discovery. Ongoing research investigates adaptive, incentive-aware ranking, federated index architectures, dynamic privacy-preserving protocols, and continual learning for profile embeddings (Raskar et al., 18 Jul 2025, Guo et al., 24 Nov 2025, Xu, 4 May 2026).

Summary Table: Representative Usable Agent Discovery Architectures

System	Identity Model	Core Discovery Mechanism	Integrity/Trust Model
NANDA	URN + AgentFacts	CRDT-index, AdaptiveResolver	Ed25519, W3C VC
AgentDNS	DNS-like/URI	Root server, hybrid semantic/RAG	PKI, TLS/JWT
agent://	URI w/ Trust Root	DHT over capability-path	PASETO attestation
ANS	ansName (DID)	K8s-native registry, CRDs	Cert chain, SNARK VC
ADS (AGNTCY)	OASF JSON-LD	Kademlia DHT, 2-level mapping	Sigstore, digest proof
IoA (2511)	AgentId + profile	PQ-augmented embedding index	Signed credentials
GRAIL	Taxo + tag/usage	SLM tag, ANN, MaxSim re-ranking	N/A (focus is retrieval)

This table is directly derived from cited system descriptions; see referenced papers for detailed API schemas, performance metrics, and implementation variants.

---

Usable agent discovery, as specified and empirically validated in recent literature, constitutes the foundational substrate for emerging agentic ecosystems. The central principles—cryptographically verifiable identity/capabilities, dynamic and policy-aware endpoint resolution, federated and scalable registry protocols, context-sensitive semantic search, and multi-criteria negotiation—enable robust, future-proof, and autonomous agent interoperability at Internet scale (Zinky et al., 5 Aug 2025, Raskar et al., 18 Jul 2025, Rodriguez, 21 Jan 2026, Muscariello et al., 23 Sep 2025, Mittal et al., 29 Apr 2026, Dazzi et al., 25 Apr 2026, Xu, 4 May 2026, Cui et al., 28 May 2025, Guo et al., 24 Nov 2025).