Overview
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. This protocol maps agent skills to directory record identifiers and maintains a list of directory servers currently hosting those records. Directory records are identified by globally unique names that are routable within a DHT (Distributed Hash Table) to locate peer directory servers. Similarly, the skill taxonomy is routable in the DHT to map skillsets to records that announce those skills.
The Agent Directory leverages the OASF to describe AI agents and provides a set of APIs and tools to build, store, publish and discover AI agents across the network by their attributes and constraints. Directory also leverages the CSIT for continuous system integration and testing across different versions, environments, and features.
Each directory record must include skills from a defined taxonomy, as specified in the Taxonomy of AI Agent Skills from the OASF. While all record data is modeled using the OASF, only skills are leveraged for content routing in the distributed network of directory servers. The ADS specification is under active development and is published as an Internet Draft at ADS Spec. The source code is available in the ADS Spec sources. The current reference implementation, written in Go, provides server and client nodes with gRPC and protocol buffer interfaces. The directory record storage is built on ORAS (OCI Registry As Storage), while data distribution uses the zot OCI server implementation.
Features
ADS enables several key capabilities for the agentic AI ecosystem:
- Capability-Based Discovery: Agents publish structured metadata describing their functional characteristics as described by the OASF. The system organizes this information using hierarchical 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.
- 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.
Naming
In distributed systems, having a reliable and collision-resistant naming scheme is crucial. The agent directory uses cryptographic hashes to generate globally unique identifiers for data records. ADS leverages Content Identifiers for naming directory records. CIDs provide a self-describing, content-addressed naming scheme that ensures data integrity and immutability.
Content Routing
ADS implements capability-based record discovery through a hierarchical skill taxonomy. This architecture enables:
- Capability Announcement:
- Multi-agent systems can publish their capabilities by encoding them as skill taxonomies.
- Each record contains metadata describing the agent's functional capabilities.
- Skills are structured in a hierarchical format for efficient matching.
- Discovery Process: The system performs a two-phase discovery operation:
- Matches queried capabilities against the skill taxonomy to determine records by their identifier.
- Identifies the server nodes storing relevant records.
- Distributed Resolution: Local nodes execute targeted retrievals based on:
- Skill matching results: Evaluates capability requirements.
- Server location information: Determines optimal data sources.
ADS uses libp2p Kad-DHT for server and content discovery.
Distributed Object Storage
ADS differs from block storage systems like IPFS in its approach to distributed object storage. The differences are described in the following sections.
Simplified Content Retrieval
- ADS directly stores complete records rather than splitting them into blocks.
- No special optimizations needed for retrieving content from multiple sources.
- Records are retrieved as complete units using standard OCI protocols.
OCI Integration
ADS leverages the OCI distribution specification for content storage and retrieval:
- Records are stored and transferred using OCI artifacts.
- Any OCI distribution-compliant server can participate in the network.
- Servers retrieve records directly from each other using standard OCI protocols.
While ADS uses zot as its reference OCI server implementation, the system works with any server that implements the OCI distribution specification.
Flow Diagrams
sequenceDiagram
participant User
participant DHT
participant ServerA
participant ServerB
participant ServerC
Note over ServerA,ServerC: Publication Phase
ServerA->>ServerA: Generate record CID
ServerA->>ServerA: Extract skills from record
ServerA->>ServerA: Store record locally
ServerA->>DHT: Announce CID + skills
ServerB->>ServerB: Generate record CID
ServerB->>ServerB: Extract skills from record
ServerB->>ServerB: Store record locally
ServerB->>DHT: Announce CID + skills
DHT->>DHT: Update routing tables<br/>(skills→CIDs→servers)
Note over User,ServerC: Discovery Phase
User->>DHT: Query by skills
DHT->>DHT: Search routing tables
DHT->>User: Return matching CIDs<br/>+ server addresses
User->>User: Select records
User->>ServerA: Download record 1
User->>ServerB: Download record 2