Independent Submission L. Muscariello Internet-Draft R. Polic Intended status: Informational Cisco Expires: 13 February 2026 12 August 2025 Agent Directory Service draft-agntcy-ads-latest Abstract The Agent Directory Service (ADS) is a distributed directory service designed to store metadata for AI agent applications. This metadata, stored as directory records, enables the discovery of agent applications with specific skills for solving various problems. The implementation features distributed directories that interconnect through a content-routing protocol. About This Document This note is to be removed before publishing as an RFC. The latest revision of this draft can be found at https://spec.dir.agncty.org. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-agntcy-ads/. Source for this draft and an issue tracker can be found at https://github.com/agntcy/dir. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at https://datatracker.ietf.org/drafts/current/. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on 13 February 2026. Copyright Notice Copyright (c) 2025 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/ license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Table of Contents 1. Conventions and Definitions 2. Introduction 2.1. Core Capabilities 2.2. Architectural Foundation 3. Storage Architecture 3.1. Content-Addressed Storage 3.2. ORAS Integration 3.2.1. Standards Compliance 3.2.2. Artifact Organization 3.2.3. Multi-Registry Federation 4. MAS Data Discovery 4.1. Skill Taxonomy 4.1.1. The Challenge of Capability Search 4.1.2. Taxonomy-Driven Search Optimization 4.2. Two-Level Mapping Architecture 4.3. Skill Taxonomy for Search Optimization 4.4. Additional Taxonomies 4.4.1. Domain Taxonomy 4.4.2. Feature Taxonomy 4.4.3. Multi-Dimensional Search 4.4.4. Skills-to-CID Mapping 4.4.5. CID-to-PeerID Mapping 4.5. DHT-Based Discovery Process 4.5.1. Skill Registration 4.5.2. Discovery Query Resolution 4.5.3. Additional Tanomoxies 4.6. Content Distribution via OCI Protocol 4.6.1. Peer-to-Peer Synchronization 4.6.2. Distribution Strategies 4.7. Agent Directory Record Examples 4.8. Security Model 4.8.1. Cryptographic Integrity 4.8.2. Content Provenance and Digital Signatures 4.8.3. Trust Boundaries and Isolation 4.8.4. Threat Mitigation 4.8.5. Access Control 4.8.6. Trust Boundaries 4.9. Performance Optimizations 4.9.1. Bandwidth Optimization 4.9.2. Scalability Architecture 5. IANA Considerations 6. References 6.1. Normative References 6.2. Informative References Authors' Addresses 1. Conventions and Definitions The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here. 2. Introduction Multi-Agent Systems (MAS) represent a new paradigm in distributed computing where software components leverage Large Language Models (LLMs) to perform specialized tasks and solve complex problems through collaborative intelligence. These systems combine LLMs with contextual knowledge and tool-calling capabilities, often abstracted through Model Context Protocol (MCP) servers, enabling dynamic workflows that adapt based on stored state and environmental conditions. The diversity and complexity of MAS architectures present unique challenges for discovery and composition. As the ecosystem of AI agents expands, developers need efficient mechanisms to: * *Discover compatible agents* with specific skills and capabilities * *Evaluate performance characteristics* including cost, latency, and resource requirements * *Compose multi-agent workflows* by linking agents with complementary capabilities * *Verify claims* about agent performance and reliability * *Track versioning and dependencies* between agent components The Agent Directory Service (ADS) addresses these challenges by providing a distributed directory infrastructure specifically designed for the agentic AI ecosystem. Rather than attempting to formally define MAS architectures, which would constrain the creative composition patterns emerging in this rapidly evolving field—ADS focuses on providing flexible metadata storage and discovery mechanisms. 2.1. Core Capabilities ADS enables several key capabilities for the agentic AI ecosystem: *Capability-Based Discovery*: Agents publish structured metadata describing their functional abilities, costs, and performance characteristics. The system organizes this information using hierarchical skill taxonomies, enabling efficient matching of capabilities to requirements. *Verifiable Claims*: While agent capabilities are often subjectively evaluated, ADS provides cryptographic mechanisms for data integrity and provenance tracking. This allows users to make informed decisions about agent selection while enabling reputation systems to emerge organically. *Semantic Linkage*: Components can be securely linked to create various relationships like version histories for evolutionary development, collaborative partnerships where complementary skills solve complex problems, and dependency chains for composite agent workflows. *Distributed Architecture*: Built on proven distributed systems principles, ADS uses content-addressing for global uniqueness and implements distributed hash tables [DHT] for scalable content discovery across decentralized networks. 2.2. Architectural Foundation The system leverages the Open Agentic Schema Framework (OASF) to model agent information in a structured, extensible format. OASF enables rich queries such as "What agents can solve problem A?" or "What combination of skills and costs optimizes for task B?" This schema-driven approach supports both objective metrics (token consumption, GPU requirements) and subjective evaluations (user ratings, task completion quality). Agent records are organized using modular extensions—reusable components like MCP server definitions, prompt-based agents, and evaluation metrics. This modular approach facilitates composition and reuse across different MAS architectures while maintaining flexibility for innovative use cases. The underlying storage layer integrates with OCI (Open Container Initiative) [OCI.Image] standards, enabling interoperability with existing container ecosystems and leveraging mature tooling for content distribution and verification. This document details the technical architecture of ADS, covering the record storage layer, security model, distributed data discovery mechanisms, and data distribution protocols between storage nodes. 3. Storage Architecture ADS implements a decentralized storage architecture built on OCI (Open Container Initiative) registries using ORAS (OCI Registry as Storage) as the foundational object storage layer. This design choice enables the system to leverage mature, standardized container registry infrastructure while achieving the speed, scalability, and security requirements of a distributed agent directory. 3.1. Content-Addressed Storage The storage architecture centers on globally unique Content Identifiers (CID) that provide several critical properties for a distributed agent directory: *Immutability*: Content identifiers are cryptographically derived from the data they represent, ensuring that any modification results in a different identifier. This property is essential for maintaining data integrity in agent records and enabling verifiable claims about agent capabilities. *Deduplication*: Identical content automatically receives the same identifier across all nodes in the network, eliminating storage redundancy and reducing bandwidth requirements when the same agent components are referenced by multiple systems. *Verifiability*: Any node can independently verify that received content matches its identifier, providing built-in protection against data corruption or tampering during transmission. *Location Independence*: Content can be retrieved from any node that possesses it, as the identifier serves as a universal pointer that abstracts away physical storage locations. 3.2. ORAS Integration ORAS provides a standardized interface for treating OCI registries as general-purpose object storage, offering several advantages for ADS: 3.2.1. Standards Compliance By building on OCI specifications, ADS inherits compatibility with the extensive ecosystem of container registry tools, security scanners, and management platforms. This includes: * *Authentication and authorization* mechanisms already deployed in enterprise environments * *Content signing and verification* through tools like Notary and cosign * *Vulnerability scanning* capabilities that can be extended to agent security assessments * *Content delivery networks* optimized for OCI artifact distribution 3.2.2. Artifact Organization Agent records are stored as OCI artifacts with a structured organization. Multiple records can be stored under the same OCI name and tag, with each record uniquely identified by its content- addressed SHA256 digest: null_repo/records/ ├── skills/ │ ├── nlp/ │ │ ├── sentiment-analysis:v1.0.0@sha256:abc123... # BERT │ │ ├── sentiment-analysis:v1.0.0@sha256:def456... # RoBERTa │ │ ├── sentiment-analysis:v1.0.0@sha256:ghi789... # DistilBERT │ │ ├── text-classification:v2.0.0@sha256:abc123... # Same BERT │ │ └── emotion-detection:v1.5.0@sha256:abc123... # Same BERT │ ├── vision/ │ │ ├── object-detection:v2.1.0@sha256:jkl012... # YOLO │ │ ├── object-detection:v2.1.0@sha256:mno345... # R-CNN │ │ └── scene-understanding:v1.0.0@sha256:jkl012... # Same YOLO │ └── reasoning/ │ └── mathematical:v1.5.0@sha256:pqr678... ├── evaluations/ │ ├── performance-metrics:latest@sha256:stu901... │ └── benchmark-results:v1.0.0@sha256:vwx234... └── compositions/ ├── security-analyst:v3.0.0@sha256:yza567... └── research-assistant:v2.2.0@sha256:bcd890... This naming scheme demonstrates that the same content identifier can belong to multiple skills, reflecting the reality that many AI agents are multi-capable. For example, the BERT-based agent (sha256:abc123...) appears under multiple skill categories: nlp/ sentiment-analysis, nlp/text-classification, and nlp/emotion- detection, representing different capabilities of the same underlying agent implementation. Similarly, the YOLO vision model (sha256:jkl012...) provides both object-detection and scene- understanding capabilities. This cross-referencing approach allows agents to be discovered through any of their supported capabilities while maintaining unique addressability through content identifiers. Each skill category can have its own versioning and metadata, enabling fine-grained capability management even when multiple skills share the same underlying implementation. Each artifact contains structured metadata following OASF schemas, enabling rich queries and capability matching across all variants within a given category. 3.2.3. Multi-Registry Federation The architecture supports federation across multiple registry instances, enabling: * *Organizational boundaries*: Different organizations can maintain their own registries while participating in the global directory * *Geographic distribution*: Content can be replicated to registries closer to consumers, reducing latency * *Specialization*: Registries can focus on specific domains (e.g., medical AI agents, financial analysis tools) * *Redundancy*: Critical agent records can be replicated across multiple registries for availability 4. MAS Data Discovery ADS implements a two-level mapping system that enables efficient discovery of Multi-Agent System components through a distributed hash table [DHT] architecture. This approach separates capability-based discovery from content location, providing both scalability and flexibility in agent retrieval. 4.1. Skill Taxonomy Effective agent discovery in multi-agent systems requires sophisticated organization of capabilities and skills. ADS employs a hierarchical skill taxonomy that serves as the foundation for efficient search and discovery operations across the distributed network. 4.1.1. The Challenge of Capability Search Traditional keyword-based search approaches face significant limitations when applied to agent discovery: *Vocabulary Fragmentation*: Different publishers may describe similar capabilities using varying terminology. For example, "sentiment analysis," "opinion mining," and "emotional classification" may all refer to similar agent capabilities, leading to search results that miss relevant agents due to terminology mismatches. *Scale Complexity*: As the number of agents in the ecosystem grows, exhaustive search across all records becomes computationally prohibitive. Without structured organization, every query potentially requires examining every agent record, leading to poor performance characteristics. *Semantic Relationships*: Many agent capabilities have natural hierarchical relationships that flat keyword systems cannot capture. An agent capable of "named entity recognition" is inherently relevant to searches for broader "text analysis" capabilities, but keyword matching alone cannot establish these connections. 4.1.2. Taxonomy-Driven Search Optimization ADS addresses these challenges through a structured hierarchical taxonomy that provides several critical optimization benefits: *Search Space Partitioning*: The taxonomy enables efficient partitioning of the search space. When processing a query for "computer vision" capabilities, the system can immediately focus on the relevant taxonomy branch, eliminating the need to examine agents in unrelated categories like natural language processing or mathematical reasoning. *Index Structure Optimization*: The hierarchical organization allows the distributed hash table to create specialized indices for different taxonomy branches. Rather than maintaining a single massive index, the DHT can distribute indexing responsibility across nodes, with each node specializing in specific capability domains. *Query Semantic Expansion*: The taxonomy enables intelligent query expansion where searches automatically include semantically related subcategories. A search for "text analysis" can transparently include results from "sentiment analysis," "entity extraction," and "text classification" without requiring users to explicitly enumerate all relevant subcategories. *Standardized Vocabulary*: By providing a canonical taxonomy, ADS reduces terminology fragmentation. Publishers are encouraged to tag their agents using standardized skill categories, improving search precision and recall across the ecosystem. 4.2. Two-Level Mapping Architecture The discovery system operates through two distinct mapping layers: 4.3. Skill Taxonomy for Search Optimization ADS employs a hierarchical skill taxonomy to optimize search performance and enable efficient capability-based discovery. Taxonomies provide several critical advantages for agent discovery systems: *Search Space Reduction*: Rather than performing exhaustive searches across all agent records, taxonomies allow the system to quickly narrow the search space to relevant categories. When a user queries for "natural language processing" capabilities, the system can immediately identify the subset of agents tagged with NLP skills without examining agents focused on computer vision or mathematical reasoning. *Hierarchical Organization*: Skills are organized in a tree-like structure that reflects natural relationships between capabilities. For example: Natural Language Processing ├── Text Analysis │ ├── Sentiment Analysis │ ├── Named Entity Recognition │ └── Text Classification ├── Language Generation │ ├── Text Summarization │ ├── Content Creation │ └── Translation └── Conversational AI ├── Dialogue Management ├── Intent Recognition └── Response Generation This hierarchy enables both specific queries ("sentiment analysis agents") and broader capability searches ("all natural language processing agents") while maintaining efficient indexing structures. *Query Expansion and Refinement*: Taxonomies support automatic query expansion where searches for parent categories can include relevant child categories. A query for "text analysis" can automatically include agents tagged with "sentiment analysis," "named entity recognition," and "text classification" without requiring users to know all specific subcategories. *Semantic Consistency*: Standardized taxonomies reduce ambiguity and improve search precision by providing consistent terminology across the ecosystem. This prevents fragmentation where similar capabilities are described using different terms by different publishers. *Scalable Indexing*: The hierarchical structure enables efficient distributed indexing where different DHT nodes can specialize in specific taxonomy branches, distributing both storage load and query processing across the network. 4.4. Additional Taxonomies While skills form the primary taxonomy for capability-based discovery, ADS supports multiple parallel taxonomies to enable rich, multi-dimensional agent classification and search. 4.4.1. Domain Taxonomy The domain taxonomy organizes agents by their application domains, representing the broader problem spaces or industries where agents are designed to operate: Application Domains ├── Networking │ ├── Network Configuration │ ├── Traffic Analysis │ └── Protocol Implementation ├── Security │ ├── Threat Detection │ ├── Vulnerability Assessment │ └── Access Control ├── Software Development │ ├── Code Generation │ ├── Testing Automation │ └── Documentation ├── Finance and Business │ ├── Risk Analysis │ ├── Market Research │ └── Process Automation └── Healthcare ├── Medical Imaging ├── Clinical Decision Support └── Drug Discovery Domain classification enables users to discover agents that are specifically tuned for their operational context, even if those agents share similar underlying skills with agents from other domains. 4.4.2. Feature Taxonomy The feature taxonomy categorizes agents by the integration frameworks and architectural patterns they support, facilitating the discovery of agents compatible with specific system architectures: Integration Features ├── MCP (Model Context Protocol) │ ├── MCP Server Implementation │ ├── MCP Client Integration │ └── MCP Tool Providers ├── A2A (Agent-to-Agent Communication) │ ├── Direct Messaging │ ├── Event-Driven Architecture │ └── Workflow Orchestration ├── Agent Evaluation │ ├── Performance Benchmarking │ ├── Quality Assessment │ └── Comparative Analysis └── Observability ├── Metrics Collection ├── Distributed Tracing └── Health Monitoring 4.4.3. Multi-Dimensional Search The parallel taxonomy system enables sophisticated multi-dimensional queries that combine criteria across different classification axes. *All searches must include at least one skill criterion as the mandatory foundation*, with domain and feature taxonomies providing additional filtering dimensions: * *Skill + Domain*: "Find natural language processing agents specialized for healthcare applications" * *Skill + Feature*: "Discover computer vision agents that support MCP integration" * *Skill + Feature + Domain*: "Locate natural language processing agents with observability features for manufacturing applications" *Skills as Search Foundation*: The skills taxonomy serves as the primary index structure in the DHT, making skill-based criteria mandatory for efficient query resolution. This design ensures that: * *Query Performance*: All searches leverage the optimized skills- to-CID mapping as the starting point, providing consistent performance characteristics * *Result Relevance*: Domain and feature filters are applied to skill-based result sets, ensuring functional capability remains the core selection criterion * *Index Efficiency*: The DHT can optimize storage and lookup patterns around the skills taxonomy while supporting supplementary filtering through domains and features Domain-only or feature-only queries are not supported, as they would bypass the primary indexing structure and provide results that may not have the functional capabilities required by the requesting system. This multi-taxonomic approach provides the flexibility to support diverse use cases while maintaining efficient indexing and search performance across all dimensions. 4.4.4. Skills-to-CID Mapping The first level maps agent capabilities and skills to their corresponding Content Identifiers (CID): Skills Index: "natural_language_processing" → ["sha256:abc123...", "sha256:def456...", "sha256:ghi789..."] "images_computer_vision" → ["sha256:jkl012...", "sha256:mno345..."] "analytical_skills" → ["sha256:pqr678...", "sha256:abc123..."] "multi_modal" → ["sha256:stu901...", "sha256:vwx234..."] This mapping enables queries such as "find all agents capable of natural language processing" to quickly resolve to a set of content identifiers without needing to search through individual agent records. 4.4.5. CID-to-PeerID Mapping The second level maps Content Identifiers to the Peer IDs of nodes that store the corresponding agent records: Content Location Index: "sha256:abc123..." → ["12D3KooWBhvxmvKvTYGJvXjnGBp7Ybr9WyoXkZvFnRtC4aBcDeFg", "12D3KooWXyZrUvHzPqKvTYGJvXjnGBp7Ybr9WyoXkZvFnRtC5gHi"] "sha256:jkl012..." → ["12D3KooWZaBcDeFgXyZrUvHzPqKvTYGJvXjnGBp7Ybr9WyoXkZvF", "12D3KooWGHiJkLmNoPqRsTuVwXyZ123456789AbCdEfGhIjKlMnO"] This separation allows the system to: - *Optimize for capability queries* without requiring knowledge of data locations * *Enable dynamic content replication* as peer availability changes * *Support multiple storage strategies* for the same content across different peers 4.5. DHT-Based Discovery Process The Distributed Hash Table stores and maintains both mapping layers across the network: ADS uses [Kad-DHT] [DHT] for server and content discovery by using the libp2p implementation that constitutes the IPFS core DHT [libp2p-kad-dht]. +----------------+ | DHT Node | | Content Index | +----------------+ ^ | +----------------+-----------------+ | | | +-------v------+ +------v-------+ +-----v--------+ | Server Node A | | Server Node B| | Server Node C| | Content X | | Content Y | | Content Z | +-------+------+ +------+-------+ +------+-------+ | | | | Content Exchange | +---------------+------------------+ | Content Replication Flow: 1. Servers register content with DHT 2. DHT maintains content-to-server mappings 3. Servers query DHT to locate content 4. DHT returns list of servers hosting content 5. Servers download content from peers 4.5.1. Skill Registration When agents are published to the network: 1. *Capability Extraction*: The system parses OASF records to extract skills, domains, and capabilities 2. *DHT Updates*: Skills-to-CID mappings are distributed across DHT nodes using consistent hashing 3. *Location Registration*: Peer nodes register themselves as providers for specific CIDs 4.5.2. Discovery Query Resolution Agent discovery follows a three-phase process: 1. *Capability Resolution*: Query "agents with skill X" resolves to a list of relevant CIDs via DHT lookup 2. *Location Resolution*: For each discovered CID, query DHT to find peer nodes storing the content 3. *Result Aggregation*: Combine capability matches with location information to produce actionable discovery results Discovery Flow: Query: "natural_language_processing" AND "finance_and_business" ↓ Phase 1: Skills → CIDs DHT["natural_language_processing"] → ["sha256:abc123...", "sha256:def456..."] DHT["finance_and_business"] → ["sha256:abc123...", "sha256:yza567..."] Intersection → ["sha256:abc123..."] ↓ Phase 2: CIDs → Peer IDs DHT["sha256:abc123..."] → ["12D3KooW...", "12D3KooX..."] ↓ Result: Agent sha256:abc123... available from peers 12D3KooW... and 12D3KooX... 4.5.3. Additional Tanomoxies 4.6. Content Distribution via OCI Protocol Once discovery identifies the relevant CIDs and their hosting peers, the actual agent records are retrieved using the OCI distribution protocol: 4.6.1. Peer-to-Peer Synchronization The discovered list of CIDs enables efficient content synchronization [OCI.Distribution] between peers: 1. *Content Negotiation*: Requesting peer queries hosting peers for available agent records 2. *OCI Pull Operations*: Standard OCI registry pull commands retrieve agent artifacts and metadata 3. *Incremental Sync*: Only missing or updated content is transferred, reducing bandwidth requirements 4. *Verification*: Content integrity is verified through cryptographic hash validation during transfer 4.6.2. Distribution Strategies The system supports multiple distribution patterns, each with distinct trade-offs and operational considerations: *On-Demand Retrieval*: Records are pulled from remote peers only when specifically requested, minimizing local storage requirements. _Trade-offs_: While this approach minimizes storage costs and ensures fresh content, it introduces several challenges: * *Query Latency*: Each request requires network round-trips to locate and retrieve content, increasing response times * *Network Cost*: Spurious or exploratory requests generate unnecessary network traffic and computational overhead * *DoS Vulnerability*: The system becomes susceptible to denial-of- service attacks where malicious actors can trigger expensive content retrieval operations by flooding the network with requests for non-existent or rarely-accessed agents * *Scalability Limits*: Performance degrades as the network grows due to increased query coordination overhead *Proactive Caching*: Popular or frequently accessed agents are automatically replicated to improve query response times. _Trade-offs_: This strategy offers significant scalability benefits but requires sophisticated management: * *Performance Gains*: Dramatic reduction in query latency for popular content, enabling sub-second response times * *Scalability*: Can handle high query volumes efficiently once popular content is cached locally * *Popularity Measurement*: Requires implementing metrics collection and analysis to identify which agents warrant caching. This includes tracking query frequencies, download patterns, and usage statistics across the network * *Storage Requirements*: Needs sufficient local storage capacity to maintain cached copies of popular records * *Cache Management*: Must implement cache eviction policies, freshness validation, and synchronization mechanisms * *Administrator Oversight*: Proactive caching policies must be configured and monitored by agent directory node administrators to balance storage costs with performance benefits *Strategic Replication*: Critical agents can be replicated across multiple peers to ensure high availability and reduce single points of failure. _Trade-offs_: This approach addresses availability concerns but introduces subjective complexity: * *High Availability*: Ensures critical agents remain accessible even during peer failures or network partitions * *Reduced Single Points of Failure*: Distributes risk across multiple storage locations * *Subjective Criticality*: The definition of "critical" or "useful" agents varies significantly between users, organizations, and use cases. What constitutes strategic value for financial services may be irrelevant for manufacturing applications * *Administrative Burden*: Requires agent directory node administrators to make strategic decisions about which agents warrant replication, considering factors like: - Organizational priorities and business requirements - Compliance and regulatory considerations - Cost-benefit analysis of storage versus availability - Community consensus on agent importance * *Resource Allocation*: Strategic replication consumes storage resources that could otherwise be used for proactive caching of popular content *Administrative Management*: Both Proactive Caching and Strategic Replication require active management by agent directory node administrators. Administrators must: * Configure caching policies based on local network characteristics and storage capacity * Monitor popularity metrics and adjust caching strategies accordingly * Define strategic replication criteria aligned with organizational objectives * Balance resource allocation between different distribution strategies * Implement governance policies for content lifecycle management This architecture provides a scalable foundation for MAS data discovery that can efficiently handle large networks of distributed agents while maintaining low latency for capability-based queries. 4.7. Agent Directory Record Examples The following examples illustrate the structure of OASF-compliant agent records stored in the directory: { "content_id": "sha256:abc123...", "record": { "name": "BERT Sentiment Analyzer", "version": "1.0.0", "schema_version": "0.2.0", "description": "Multi-capability NLP agent providing sentiment analysis, text classification, and emotion detection", "skills": ["natural_language_processing"], "domains": ["finance_and_business", "trust_and_safety"], "capabilities": { "threads": true, "interrupt_support": false, "callbacks": true, "streaming": ["text", "json"] } }, "performance_metrics": { "tokens_per_second": 1000, "gpu_memory_mb": 4096, "latency_p99_ms": 150, "accuracy_score": 0.94 }, "evaluation_data": { "overall_rating": 4.2, "cost_per_million_tokens": 2.50 }, "registries": [ "registry.example.com", "hub.agents.org" ], "last_updated": "2025-08-07T10:30:00Z" } { "content_id": "sha256:jkl012...", "record": { "name": "YOLO Vision Agent", "version": "2.1.0", "schema_version": "0.2.0", "description": "Computer vision agent for object detection and scene understanding", "skills": ["images_computer_vision"], "domains": ["transportation", "industrial_manufacturing"], "capabilities": { "threads": false, "interrupt_support": true, "callbacks": false, "streaming": ["image", "json"] } }, "performance_metrics": { "inference_fps": 30, "gpu_memory_mb": 8192, "detection_accuracy_map": 0.89, "processing_latency_ms": 33 }, "evaluation_data": { "overall_rating": 4.7, "cost_per_image": 0.05 }, "registries": [ "vision.agents.com", "registry.example.com" ], "last_updated": "2025-08-07T14:20:00Z" } { "content_id": "sha256:pqr678...", "record": { "name": "Mathematical Reasoning Agent", "version": "1.5.0", "schema_version": "0.2.0", "description": "Agent specialized in mathematical problem solving and analytical reasoning", "skills": ["analytical_skills", "tabular_text"], "domains": ["education", "finance_and_business"], "capabilities": { "threads": true, "interrupt_support": true, "callbacks": true, "streaming": ["text", "latex"] } }, "performance_metrics": { "problems_per_minute": 12, "cpu_cores": 4, "memory_mb": 2048, "accuracy_on_gsm8k": 0.87 }, "evaluation_data": { "overall_rating": 4.5, "cost_per_problem": 0.10 }, "registries": [ "math.agents.edu", "registry.example.com" ], "last_updated": "2025-08-07T16:45:00Z" } These examples demonstrate how the DHT indexing system extracts skills and domains from agent records to populate the Skills-to-CID mappings, enabling efficient capability-based discovery across the distributed network. 4.8. Security Model The OCI-based architecture provides multiple layers of security that address the unique challenges of distributed agent directories: 4.8.1. Cryptographic Integrity ADS leverages the OCI layer's built-in cryptographic mechanisms to ensure data integrity: *Automatic Hash Computation*: The OCI registry layer automatically computes SHA-256 hash digests for all stored artifacts. These content identifiers are generated transparently during the push operation, ensuring that every agent record has a cryptographically verifiable fingerprint without additional overhead. *Tamper Detection*: Content addressing ensures immediate tamper detection through cryptographic hash verification. Any modification to an agent record—whether malicious or accidental—results in a different hash digest, making unauthorized changes immediately detectable during retrieval operations. *End-to-End Verification*: Clients can independently verify that received content matches its advertised identifier, providing built- in protection against data corruption during transmission or storage without trusting intermediate network components. 4.8.2. Content Provenance and Digital Signatures ADS integrates with Sigstore, a security framework for OCI storage, to provide comprehensive content provenance and authenticity guarantees: *Sigstore Integration*: The system leverages Sigstore's security framework to provide verifiable proof of when and by whom agent records were signed. This creates an immutable audit trail that cannot be retroactively modified, enabling forensic analysis of agent deployment history. *Keyless Signing*: Sigstore's keyless signing approach eliminates the complexity and security risks associated with long-lived cryptographic keys: * *Identity-Based Authentication*: Uses OpenID Connect (OIDC) [OpenID.Auth] tokens from trusted identity providers (GitHub, Google, Microsoft) to authenticate publishers at signing time * *Short-Lived Certificates*: Issues ephemeral signing certificates valid only for minutes, reducing the window of potential key compromise * *Automatic Key Rotation*: Eliminates the need for manual key management, distribution, and rotation procedures * *Scalable Trust*: Publishers don't need to maintain or distribute public keys, making the system accessible to individual developers and large organizations alike *Transparency and Verification*: All signatures are stored directly in OCI storage alongside the agent artifacts and public keys, providing: * *Public Auditability*: Anyone can verify the signing history of agent records stored in accessible registries * *Non-Repudiation*: Publishers cannot deny having signed records that are cryptographically linked to their identity * *Supply Chain Security*: Enables detection of compromised or unauthorized agent publications 4.8.3. Trust Boundaries and Isolation *Organizational Isolation*: Separate registries maintain security boundaries between different organizations, allowing each entity to control their own agent ecosystem while still participating in the broader federated network. *Content Verification*: Nodes can validate artifact integrity and signature authenticity without trusting transport layers or intermediate storage systems. This zero-trust approach ensures security even when using untrusted storage infrastructure. *Reputation Systems*: The cryptographic foundation enables the development of reputation systems based on verifiable evidence rather than subjective claims. Publishers with consistent signing practices and high-quality agents can build measurable trust over time. 4.8.4. Threat Mitigation The security model addresses several key threats to distributed agent directories: *Supply Chain Attacks*: Sigstore integration and transparency logs make it difficult for attackers to inject malicious agents without detection, as all publications are cryptographically signed and publicly auditable. *Data Integrity Attacks*: Automatic hash verification prevents tampering with agent records during storage or transmission, ensuring users receive authentic content. *Identity Spoofing*: OIDC-based keyless signing prevents attackers from impersonating legitimate publishers without compromising their identity provider credentials. *Availability Attacks*: The distributed nature of the system, combined with content replication across multiple registries, provides resilience against denial-of-service attacks targeting individual nodes. 4.8.5. Access Control * *Registry-level permissions* control who can publish and retrieve agent records * *Fine-grained policies* can restrict access to specific agent categories or capability types * *Audit trails* leverage existing registry logging capabilities to track access patterns 4.8.6. Trust Boundaries * *Organizational isolation* through separate registries maintains security boundaries * *Content verification* allows nodes to validate artifact integrity without trusting transport layers * *Reputation systems* can build on cryptographic proofs of past agent performance 4.9. Performance Optimizations The architecture incorporates several optimizations for the specific requirements of agent discovery: 4.9.1. Bandwidth Optimization * *Incremental updates* use OCI layer semantics to transmit only changed portions of agent records * *Content compression* reduces storage and transmission costs for large agent definitions * *Selective replication* based on query patterns minimizes unnecessary data transfer 4.9.2. Scalability Architecture The system scales horizontally through several mechanisms: * *Registry sharding* distributes storage load across multiple OCI registry instances * *Index partitioning* in the DHT allows query load to scale with the number of participating nodes * *Lazy loading* defers retrieval of detailed agent specifications until actually needed This architecture provides a robust foundation for a decentralized agent directory that can scale to support the growing ecosystem of AI agents while maintaining the security and reliability requirements of production systems. 5. IANA Considerations This document has no IANA actions. 6. References 6.1. Normative References [OCI.Distribution] Initiative, O. C., "OCI Distribution Specification", n.d., . [OCI.Image] Initiative, O. C., "OCI Image Format Specification", n.d., . [OpenID.Auth] Foundation, O., "OpenID Authentication 2.0 - Final", n.d., . [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, . [RFC6920] Farrell, S., Kutscher, D., Dannewitz, C., Ohlman, B., Keranen, A., and P. Hallam-Baker, "Naming Things with Hashes", RFC 6920, DOI 10.17487/RFC6920, April 2013, . [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017, . 6.2. Informative References [DHT] Maymounkov, P. and D. Mazieres, "Kademlia: A peer-to-peer information system based on the xor metric", IPTPS '01 , 2001. [libp2p-kad-dht] Community, libp2p., "go-libp2p-kad-dht: A Kademlia DHT implementation on go-libp2p", n.d., . [Sigstore] Foundation, C. N. C., "Sigstore: A New Standard for Signing, Verifying and Protecting Software", n.d., . Authors' Addresses Luca Muscariello Cisco Email: lumuscar@cisco.com Ramiz Polic Cisco Email: rpolic@cisco.com