AgentDNS: LLM Service Discovery System

• AgentDNS is a robust naming and service discovery system that abstracts logical identities for LLM agents and tools.
• It employs semantic matching and hybrid retrieval to enable natural language queries and protocol-aware interoperability.
• It unifies authentication and billing, allowing autonomous multi-agent workflows through secure, centralized service management.

AgentDNS is a root domain naming and service discovery system tailored for LLM agents and their tool ecosystem. It is explicitly inspired by the canonical Domain Name System (DNS), yet engineered to address critical gaps in automated service discovery, cross-vendor interoperability, secure invocation, and unified billing within multi-agent architectures. AgentDNS abstracts service identities from physical endpoints, provides intent-based semantic discovery, supports protocol-aware interoperability, and centralizes authentication and economic transactions for agent interactions. The open-source implementation, designated for release at https://github.com/agentdns, is described as foundational infrastructure for autonomous, multi-agent collaboration across organizational and technological boundaries (Cui et al., 28 May 2025).

1. System Architecture and Naming Scheme

AgentDNS is architected around a central root server hub that orchestrates service registration, proxy management, semantic service search, secure metadata resolution, authentication, and unified accounting. Vendors register agent or tool services under unique hierarchical identifiers, following the schema:

agentdns://organization/category/name

This structure supports hierarchical categorization (e.g., agentdns://openai/nlp/textsummarization). Service metadata includes endpoints, supported protocols (e.g., Model Context Protocol (MCP), Agent-to-Agent Protocol (A2A)), capability descriptors, and price. Registered services are reflected in a proxy pool, which abstracts actual endpoints and manages protocol-specific routing and authentication mediation.

Agents interact with AgentDNS via discovery (natural language queries), metadata resolution (lookup and refresh of identifier-bound information), and invocation through time-bound access tokens. Unified Single-Sign-On authentication and centralized billing allow agents to invoke heterogeneous services without individual vendor credentials.

2. Semantic Service Discovery and Registration

Service registration associates logical identities with comprehensive metadata, decoupling discovery and invocation from static network coordinates. Agents may submit queries in natural language (such as "Find an agent that analyzes academic papers"), which AgentDNS resolves using hybrid retrieval strategies combining keyword matching with retrieval-augmented generation operating on agent and tool capability embeddings.

Discovery returns a ranked list of candidate services exposing identifier, protocol compatibility, endpoint, cost, and capability metadata. These results facilitate agents' automated selection and negotiation processes, enabling workflow construction with minimal manual intervention. Semantic matching is not restricted to exact name queries; broad intent-based matching is supported by the underlying vector similarity and filtering algorithms.

3. Protocol-Aware Interoperability and Secure Invocation

Agents discover, resolve, and adapt to vendor-specific communication protocols through the returned metadata (e.g., MCP, A2A, ANP). Agents invoke services solely through AgentDNS proxies using issued access tokens, which mediate authentication and enforce protocol transformations, eliminating the need for agents to manage distinct vendor API keys or custom protocol adapters.

Time-bound tokens conform to Single-Sign-On semantics across the AgentDNS federation, and only agents presenting valid credentials can access registered service proxies. The architecture thus enforces trustless authentication with transparent mediation.

4. Unified Billing and Economic Integration

AgentDNS maintains centralized tracking of usage, cost, and payment for all registered services. User agents deposit funds to a single account at AgentDNS, which meters service consumption and distributes payments to underlying vendors. This unifies otherwise fragmented economic systems, supporting both free and paid agent/tool services and enabling billing for autonomous agents without direct human management.

Granular billing records and backend settlement mechanisms elevate the agent economy by removing the need for per-vendor negotiation and manual invoice reconciliation.

5. Technical Mechanisms and DNS Analogy

AgentDNS inherits core DNS principles of decoupling logical names from physical addresses but extends these with semantic richness, protocol metadata, authentication tokens, and payment abstractions. Identifier resolution follows: $$\texttt{agentdns://}\langle\text{organization}\rangle/\langle\text{category}\rangle/\langle\text{name}\rangle$$ Metadata returned upon query includes: $$\text{protocols} \subseteq \{ \text{MCP}, \text{A2A}, \text{ANP}, \ldots \}$$ Service discovery (hybrid retrieval) is implemented by embedding both queries and service descriptions as vector representations:

• For all registered services $S_i$ with capability embedding $E_i$,
• The query $Q$ is embedded as $E_Q$,
• Similarity $sim(E_Q, E_i)$ is computed,
• Top-$k$ services are returned after filtering by intent, protocol, or pricing constraints.

Caching and periodic refresh of identifier-resolved metadata ensure currency and performance.

6. Use Cases and Autonomous Multi-agent Workflows

AgentDNS enables fully autonomous multi-agent workflows. An agent tasked with a complex objective (e.g., generating a research survey) decomposes the problem into actionable steps, submits semantic queries to AgentDNS to discover suitable agents and tools, retrieves protocol and economic metadata, adapts invocation logic according to capabilities, and leverages the unified billing and authentication system to orchestrate the entire workflow without human oversight or static endpoint configuration.

A salient case paper involves an agent building a multi-step plan, issuing intent-based queries, resolving protocol and billing details via AgentDNS, and performing end-to-end invocation and output composition across diverse vendor tools. The architecture natively decouples planning and execution logic from fixed vendor endpoints.

7. Comparison to DNS and Future Directions

AgentDNS is designed to remediate the limitations of traditional DNS, which offers only static, human-readable naming and IP address resolution. It supports hierarchical, semantic naming, endpoint and protocol metadata mapping, natural language service discovery, protocol negotiation, Single-Sign-On authentication, and unified billing. The design proposes future evolution toward decentralized or federated models, enhanced security (homomorphic encryption, differential privacy), and reputation-based economic systems.

Feature	Traditional DNS	AgentDNS
Naming	Hostname	Semantically-rich, hierarchical ID
Resolution	IP address	Endpoints + protocols + metadata
Discovery	Exact	Semantic, natural language search
Authentication	None	Unified token-based SSO
Billing	None	Unified, pay-as-you-go, multi-vendor
Protocol Info	Not supported	Protocol negotiation supported

8. Implications for Multi-Agent Ecosystems

AgentDNS operationalizes critical infrastructure for LLM agents by enabling autonomous, intent-driven service discovery, adaptive protocol negotiation, trustless authentication, and economic abstraction across disparate vendor domains. The system aligns with trends toward scalable, agent-centric automation and interoperability, directly supporting research and production deployments in agentic AI and tool ecosystems.

The open-source implementation and planned decentralization facilitate further community-driven development, standardization, and analysis for collaborative LLM systems. AgentDNS thus constitutes the backbone for scalable, secure, and vendor-agnostic multi-agent collaboration in contemporary and emerging computational settings (Cui et al., 28 May 2025).