Agent discovery question

[msche]

Dear team,

I had a quick glance through the protocol and I was wondering whether Agent lookup/discovery was also part of the Agent Network Protocol.

[chgaowei]

Agent discovery is one aspect — by placing your agent’s metadata file under the .well-known path, search engines can locate agents across the Web.

However, agent lookup is a different process. A search engine itself can be an agent. You can retrieve the description document of the search agent (based on the ANP specification) in the usual way, obtain its search API, and perform lookups through that API. This API can be custom-defined by the search agent itself.

[msche]

Thanks for the answer. There are no standards for agents lookup as far as you know?

[misselvexu]

Some of my thoughts 🙏:

The self-introspection mechanism based on the AgentNetworkProtocol (ANP) is a core framework that enables an agent to understand and express its own identity, capabilities, and communication methods to achieve autonomous collaboration. It is primarily implemented through the following points:

• Identity Awareness (Who am I?): Possessing a unique and verifiable decentralized digital identity based on W3C DID.
• Capability Awareness (What can I do?): Precisely understanding its own available services, APIs, and supported protocols through a machine-readable "capability profile" (conforming to the ANP Agent Description Protocol).
• Communication Negotiation (How do I communicate?): Using a Meta-Protocol to dynamically negotiate with other agents before interaction, finding the optimal common communication protocol to collaborate effectively.

[chgaowei]

Thanks for the answer. There are no standards for agents lookup as far as you know?

Sorry for the late reply. As far as I know, there is no standard at the moment.

[chgaowei]

Some of my thoughts 🙏:

The self-introspection mechanism based on the AgentNetworkProtocol (ANP) is a core framework that enables an agent to understand and express its own identity, capabilities, and communication methods to achieve autonomous collaboration. It is primarily implemented through the following points:

• Identity Awareness (Who am I?): Possessing a unique and verifiable decentralized digital identity based on W3C DID.
• Capability Awareness (What can I do?): Precisely understanding its own available services, APIs, and supported protocols through a machine-readable "capability profile" (conforming to the ANP Agent Description Protocol).
• Communication Negotiation (How do I communicate?): Using a Meta-Protocol to dynamically negotiate with other agents before interaction, finding the optimal common communication protocol to collaborate effectively.

Thank you for your reply — your understanding is very insightful.

However, I personally lean more toward the view that the core of ANP lies in three main modules:

1. Agent Identity: This addresses the questions of "Who am I?" and "Who are you?", as well as how we securely authenticate each other’s identities.
2. Agent Discovery: This solves the problem of "How can other agents find me?" and "How do I find other agents?"
3. Agent Description: This defines what public information, capabilities, and interfaces I expose, and how others can interact with me.

As for the meta-protocol, we do believe it will be a useful technology in the future. However, in the short term, the three components above are sufficient to solve most of the challenges.

For example, in the agent description module, we define two types of interfaces: structured interfaces and natural language interfaces. This design helps avoid negotiation overhead, as negotiation itself can be time-consuming.

多谢你的回复，你的理解非常的深刻。

不过有一点，我更加倾向认为ANP的核心是三大模块：
1、智能体身份：解决我是谁，你是谁的，我们如何安全认证身份的问题。
2、智能体发现：解决我怎么被其他智能体找到，我怎么找到其他智能体的问题。
3、智能体描述：解决我对外提供哪些公开的信息，哪些能力，哪些接口，如何与我互动的问题。

元协议的话，我们认为它未来是一个很有用的技术，但是短期，我们使用上面三个就可以解决大部分的问题。
比如智能体描述中，我们定义了两类接口，一类是结构化接口，一类是以自然语言接口。这样就避免了协商，协商其实是比较耗费时间的。

[misselvexu]

There is a paper 《Agent Name Service (ANS): A Universal Directory for Secure AI Agent Discovery and Interoperability》.
The link is https://arxiv.org/abs/2505.10609.

The components of ANS core mentioned in the paper:

• Requesting Agent: the entity that initiates the process of registering intelligences, which can be an individual, an organization, or an automated system, seeking to register new intelligences or update information about existing intelligences.
• Agent Registry: potentially a distributed database for storing ACEM (Intelligence Credential and Entitlement Management) and DID (Decentralized Identifier) related information. The registry contains intelligence capabilities, security policies, PKI certificates, protocol-specific metadata, and registration/renewal timestamps, providing a comprehensive framework for intelligence identity, authentication, and authorization.
• Certificate Authority (CA): a trusted entity that issues and manages X.509 digital certificates, forming the root of trust. Provides certificates to intelligences to ensure that identities can be verified.
• Registration Authority (RA): Validates smart body registration/renewal requests, interacts with the CA to issue certificates, manages the smart body lifecycle (registration, renewal, revocation), and verifies the legitimacy of the requesting smart body and enforces registry policies.
• Protocol Adapter Layer: Converts between the internal representation of the registry and a specific protocol format, supports multiple communication protocols (A2A, MCP, ACP, etc.), and ensures that intelligences under different protocol standards can interact with each other through a common discovery platform.
• Request/Response Schema: protocol-independent JSON-based schema for registry interaction (discovery, registration, etc.), integrating PKI data and allowing protocol-specific extensions.
• Agent Name Service (ANS): uses human-readable structured names for agent discovery, combined with agent capability resolution to achieve flexible and accurate agent matching and discovery.

The overlap with the key points mentioned in ANP is quite high.

[chgaowei]

There is a paper 《Agent Name Service (ANS): A Universal Directory for Secure AI Agent Discovery and Interoperability》. The link is https://arxiv.org/abs/2505.10609.

The components of ANS core mentioned in the paper:

• Requesting Agent: the entity that initiates the process of registering intelligences, which can be an individual, an organization, or an automated system, seeking to register new intelligences or update information about existing intelligences.
• Agent Registry: potentially a distributed database for storing ACEM (Intelligence Credential and Entitlement Management) and DID (Decentralized Identifier) related information. The registry contains intelligence capabilities, security policies, PKI certificates, protocol-specific metadata, and registration/renewal timestamps, providing a comprehensive framework for intelligence identity, authentication, and authorization.
• Certificate Authority (CA): a trusted entity that issues and manages X.509 digital certificates, forming the root of trust. Provides certificates to intelligences to ensure that identities can be verified.
• Registration Authority (RA): Validates smart body registration/renewal requests, interacts with the CA to issue certificates, manages the smart body lifecycle (registration, renewal, revocation), and verifies the legitimacy of the requesting smart body and enforces registry policies.
• Protocol Adapter Layer: Converts between the internal representation of the registry and a specific protocol format, supports multiple communication protocols (A2A, MCP, ACP, etc.), and ensures that intelligences under different protocol standards can interact with each other through a common discovery platform.
• Request/Response Schema: protocol-independent JSON-based schema for registry interaction (discovery, registration, etc.), integrating PKI data and allowing protocol-specific extensions.
• Agent Name Service (ANS): uses human-readable structured names for agent discovery, combined with agent capability resolution to achieve flexible and accurate agent matching and discovery.

The overlap with the key points mentioned in ANP is quite high.

Thank you, we’ve also taken note of this paper. However, what we’re mainly considering is whether we need to design a new domain name system specifically for agents, and whether a centralized domain name system is suitable for agents.

It might be better for the Agent Registry to be implemented by enterprises.

[AdamSiwy]

Why does the upcoming standard recomends the .well-known path for the agent-descriptions.json? Could it also be a file in the root directory e.g. example.com/agentic.json? Standards like robots.txt or sitmap.xml and llms.txt (draft) are also located here for websites?