AGNTCY Agent Directory: Find and publish AI agents with our new service

At Outshift by Cisco, we know that multi-agent software and workflows are crucial to power the future of enterprise businesses. As part of the AGNTCY, we are building the open and interoperable infrastructure, also known as the Internet of Agents, to enable agent-to-agent collaboration through the entire multi-agent software lifecycle: Discover, Compose, Deploy, and Evaluate.

We will soon be releasing a free, multi-tenant SaaS version of the AGNTCY’s Agent Directory to unlock the biggest requirement that developers have when they build multi-agent software: sourcing the best agents for the job with confidence that they’re reputable, effective, and compatible with each other.

With this new service, to be released next week, we are making it simple for the community to publish their agents and keep them organized through the agent lifecycle, while standardizing the way to describe their properties and capabilities (from core attributes, skills, domains, and features, to capabilities and performance).

Developers of multi-agent software will also be able to seamlessly discover, search and source the best agents for their project, review their detailed description and pull these records for use when needed.

Read on for a technical dive into the Agent Directory service and the Open Agentic Schema Framework (OASF)—core Discovery components of AGNTCY. We’ll expand on the underlying technologies, protocols, and data models that power the Internet of Agents for agent discovery. Whether you’re a solutions architect, developer, or platform engineer, we hope this look under the hood will illuminate how these pieces can radically accelerate multi-agent software development.

What is the Agent Directory?

The AGNTCY Agent Directory allows publication, exchange, and discovery of information about artificial intelligence (AI) agents over a distributed peer-to-peer network. It leverages OASF to describe agents and provides a set of APIs and tools to build, store, publish, and discover agents across the network by their attributes and constraints. The AGNTCY Agent Directory also leverages Continuous System Integration Testing (CSIT) across different versions, environments, and features. Much more than a mere list of agents, the service is a combination of:

• An OCI-based registry that stores metadata about each agent, tracked by SHA-based digests. 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), and data distribution uses the OCI distribution spec using the zot server implementation.
• An indexing system that makes agents globally discoverable via skillsets and other attributes such as name, domain, etc. to further search in the local server.
• A cryptographic trust model that ensures the authenticity and provenance of each agent record by signing records and encrypting the digest with the owner’s private key.

The agent client SDK library brings all the capabilities to use the Directory service: add a new agent record, set property and rules (e.g.: public agent, private agent), search for agents based on various criteria, and retrieve agents' records. You can also modify the directory via the service interface or via API.

Any organization building and managing agents can download, deploy and run their instance of Agent Directory internally to streamline their development activities and agentic product lifecycle. This has several benefits:

• Centralized agent management: Developers can push new agent definitions or updates to a single, trusted repository. Platform engineers can enforce policies around version control, scanning for vulnerabilities, or validating metadata.
• Reduced duplication of effort: If multiple teams within the enterprise build similar agents (e.g., event processing or data transformation), Directory’s search feature ensures reuse and consistency.
• Better collaboration: Product owners and line-of-business leads can browse available agents without rummaging through multiple codebases. Cross-functional teams can iterate faster on prototypes and workflows, thanks to the discoverable OASF definitions.
• Compliance and security: Each agent is cryptographically verifiable, so you can confirm you’re running the correct version. Access control policies can restrict who can publish, update, or discover certain agents. The Agent Directory also exposes hooks for signature verification, so organizations can implement or integrate key management and signature workflows (e.g., using Sigstore/Cosign or other PKI solutions) to secure assets and attest provenance.

The foundation of Agent Directory: The Open Agentic Schema Framework

The Open Agentic Schema Framework (OASF) is a standardized schema system for defining and managing AI agent capabilities, interactions, and metadata. It provides a structured way to describe agent attributes, capabilities, and relationships using attribute-based taxonomies. The framework includes development tools, schema validation, and hot-reload capabilities for rapid schema development, all managed through a Taskfile-based workflow and containerized development environment. OASF serves as the foundation for interoperable AI agent systems, enabling consistent definition and discovery of agent capabilities across distributed systems.

As detailed in the OASF documentation, agent descriptions are based on a main object which includes:

• Agent Record ID: Digest to give the agent record a globally unique identifier and used to verify the record integrity via a secure hash (SHA-256)
• Agent name: A human readable name which has no indexing meaning such as “marketing strategy agent” or “Kubernetes deployment agent.”
• Version: Following semver or other versioning strategies
• Skills: List of capabilities or tasks that this agent can perform (e.g., “translate text,” “parse JSON,” “fetch data from a CRM”).
• Locators: Locators of resources to access the agent e.g. pointer to Git code, containers, SaaS endpoints).
• Extensions: Allows for additional description of agent related information for a specific purpose.

Here’s an OASF compliant representation of a simple text generation agent data model in JSON:

{
  "digest": "sha256:4e8c72f126b2e4a318911ba11b39432978d0611a56d53a2cfb6fdb42853df0e2",
  "skills": [
    "language/text-generation",
    "language/text-completion"
  ],
 "name": "gpt4-agent",
 "version": "1.0.0",
"locator": {
    "type": "source-code",
    "url": "github.com/agntcy/agents/gpt4-agent"
     }
}

A more complex example can be found here in the OASF navigator tool.

Extensions play a key role in supporting various information that correspond to specific agentic frameworks, tools, technologies or deployment patterns. For instance, in AGNTCY we are using extensions to make Discovery components (OASF and Agent Directory) useful to store information related to AGNTCY’s Agent Connect Protocol (ACP), agentic manifests (ACP is the core AGNTCY API used for agent-to-agent collaboration) or related to observability and evaluation metrics.

OASF is already rich in content. It's open and meant to be improved and expanded by the community. Here is more information if you want to contribute or propose new extensions to cover additional domain or specific industry needs for agent metadata.

👉  Did you know? We created OASF taking inspiration and branching out from OCSF, a proven and widely adopted Open Schema Framework used in the Cybersecurity domain.

👉 Important note: When you create a multi-agent application (e.g. a workflow connecting an NLP translator agent, a sentiment analysis agent, and a scheduling agent), you can bundle the resulting app into a higher-level OASF “composition” manifest. This composition itself is also versioned and published in the Directory, making complex multi-agent systems reusable building blocks for your organization or external partners.

Federating directories: The power of distributed discovery

A single Agent Directory Service is valuable, but the real magic occurs when these directories connect to each other. By leveraging a distributed hash table (DHT) and peer-to-peer (P2P) synchronization, multiple directories can: 

• Expose public agents: Mark specific agents or entire namespaces as publicly discoverable by external peers.
• Sync agent indexes: Subscribe to remote directories and pull newly published OASF manifests.
• Form a planet-scale network (AKA: The Internet of Agents): As more organizations join the Internet of Agents, the DHT ensures that any node can discover globally published agents.

The AGNTCY stack typically relies on a P2P layer (libp2p Kad-DHT) to exchange DHT entries that map:

• Agent’s announced skillset → Agent digests
• Agent digests → Directory nodes ID that host such records

This design is resilient and vendor-neutral, reflecting the same ethos that inspired decentralized version control systems like Git or distributed file systems like IPFS (InterPlanetary File System). Also note that any OCI distribution-compliant server can participate in the network. By storing references to agent data in a distributed fashion, the network:

• Improves fault tolerance—there’s no single server that, if taken down, disrupts the entire network.
• Ensures that new directories operated by various organizations can join at any time, syncing relevant agent data on-demand.

Another key design point is making search efficient at large scale across a planet-scale distributed network of directories. To address this challenge, our design, as often with DHT-based systems, uses a two-step approach:

• Initial DHT query
  The first query on the DHT searches for data models which contain the desired skillsets and return record identifiers, which are core OASF attributes, and their location via the node identifier. 

• Core OASF attributes (e.g., ID, name, skills). 

  This step efficiently returns a shortlist of agents’ data models that match the initial criteria, along with the directory nodes that host their metadata. At this stage the local node triggers an OCI data sync which pulls the selected OASF records from the best neighboring nodes.

• Comprehensive local search
  After the sync operation, all metadata is locally available, and a deeper search can be performed to select the OASF records based on a larger set of attributes such as constraints, costs and evaluation scores that might be present for some records.

This two-phase strategy balances search efficiency and thoroughness, ensuring that high-level queries remain performant while more specialized filtering takes place once the agent’s complete data has been collected locally.

AGNTCY Agent Directory: A multi-tenant service to find and publish agents

To simplify adoption, we are happy to announce that Outshift is launching a multi-tenant Agent Directory as a free SaaS. Key benefits include:

• Immediate access: Start publishing and discovering agents without deploying on-premises infrastructure.
• Community hub: The multi-tenant environment fosters collaboration and feedback, especially to improve and evolve OASF and Agent Directory components.
• On-ramp to self-hosted: When ready for production or custom policy controls, migrate to your own self-hosted Agent Directory while retaining full data portability (thanks to OASF and OCI compliance).

Our goal is to lower the barrier to entry for startups, small teams, or even large enterprises looking to evaluate the AGNTCY approach quickly.

Stay tuned for our release next week and please reach out if you have questions or want to join the AGNTCY Collective. Let’s build an Internet of Agents that works for all!