Framework for AI Agent Networks

1.1. Overview

With the development of AI agent technology, its application scenarios have been continuously expanding. From initial simple task execution to complex collaborative tasks among multiple agents, agents have demonstrated great potential in various fields. This multi-agent collaboration model can fully leverage the strengths of individual agents, improving the quality and efficiency of task execution. However, as the demand for multi-agent collaboration grows, defining standardized communication protocols among agents to achieve wide-area interconnection, cross-domain interoperability, and secure collaboration has become an urgent issue to address.¶

To meet the communication needs of AI agents and promote the widespread services of multi-agent collaboration [I-D.stephan-ai-agent-6g], it is imperative to define standardized agent communication protocols that support interconnection, interoperability, and secure scalability between agents in trust domain. In this draft we propose to use Agent Network Protocol (ANP) as a baseline for further description.¶

1.2. Scope

From the perspective of network service domain division, future agents can be simply categorized into 3 types based on their deployment locations: terminal-side agents, network-side agents, and agents outside the network. This draft mainly focuses on the communication between agents directly managed within the operator's network, i.e. the communication between the first two types of agents:¶

• Communication between different terminal-side agents registered in the same network service domain.¶

• Communication between terminal-side agents and network-side agents registered in the same network service domain.¶

• Communication between network-side agents registered in the same network service domain.¶

Furthermore, the communication between agents registered in different network domains is not within the scope of this draft.¶

1.3. Requirements Language

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.¶

3.1. Roles

The Agent communication network defines three roles:¶

• AI Agent: An automated intelligent entity that achieves a specific goal (autonomously or not) on behalf of another entity, by e.g. interacting with its environment, acquiring contextual information, reasoning, self-learning, decision-making, and executing tasks (independently or in collaboration with other AI Agents).¶

• Agent Registration Server: The server enables Agents to register their capabilities, and discover each other’s capabilities based on intent, task or other information.¶

• Agent Communication Server The server enables Agents to communicate and collaborate with each other, which provides session management and routing function.¶

3.2. Protocol Flow

+------+                           +-------------+                            +------+
|      |                           |             |<-(A)-Capability Register---|      |
|      |                           |             |                            |      |
|      |                           |    Agent    |-(B)-------Response-------->|      |
|      |-(C)-Capability Discover-->| Registration|                            |      |
|      |                           |    Server   |                            |      |
|      |<-(D)----Matched Agents----|             |                            |      |
|  AI  |                           |             |                            |  AI  |
| Agent|                           +-------------+                            | Agent|
|   A  |                                                                      |   B  |
|      |                           +-------------+                            |      |
|      |-(E)Communication Request->|             |                            |      |
|      |                           |    Agent    |-(F)Communication Request-->|      |
|      |                           |Communication|                            |      |
|      |                           |   Server    |<-(G)Communication Response-|      |
|      |<-(H)Communication Response|             |                            |      |
|      |                           +-------------+                            |      |
|      |                                                                      |      |
|      |-----------------------(E')Communication Request--------------------->|      |
|      |                                                                      |      |
|      |>----------------------(F')Communication Response---------------------|      |
+------+                                                                      +------+

The abstract flow illustrated in Figure 1 describes the interaction between the three roles and includes the following steps:¶

(A) The AI Agent B requests to register its capabilities and related attributes to Agent Registration Server.¶

(B) The Agent Registration Server authenticates the AI Agent B’s capabilities and then stores them, e.g., in its local database.¶

(C) AI Agent A initiates a capability discover request to the Agent Registration Server, the request includes the intent, task or other information.¶

(D) The Agent Registration Server matches the intent or task with the capabilities stored in its local database, and responses with matched AI Agents list to the AI Agent A.¶

Option1:¶

(E) The AI Agent A selects AI Agent B from the list and sends a communication request to AI Agent B via Agent Communication Server.¶

(F) The Agent Communication Server establishes the session and routes the message to AI Agent B.¶

(G) The AI Agent B receives the communication request and sends a response to the Agent Communication Server.¶

(H) The Agent Communication Server transfers the response to the AI Agent A.¶

Option2:¶

(E') The AI Agent A selects AI Agent B from the list and sends a communication request to AI Agent B directly.¶

(F') The AI Agent B receives the communication request and sends a response to the AI Agent A.¶

5.1. Agent Description Document Format

The Agent Description (AD) document serves as the external entry point for an agent and can be provided in either of the following formats:¶

5.2. Agent Information Interaction Mechanism

Agent description documents include the following two types of resources:¶

5.3. Security Mechanism

Security configuration in Agent Description (AD) documents is mandatory. The security definition must be activated through the security member at the agent level. This configuration constitutes the required security mechanism for agent interactions.¶

• Global Scope: When security is declared at the root level of an AD document, all resources within the document must enforce this security mechanism for access.¶

• Resource-Specific Scope: If security is defined within an individual resource, access to that resource is granted only when the specified security conditions are met.¶

• Precedence Rule: In cases where resource-level security conflicts with root-level security, the resource-level definition takes precedence.¶

5.4. Integrity Verification

To prevent malicious tampering, impersonation, or reuse of Agent Description (AD) documents, a verification mechanism Proof is incorporated into the AD document structure. The definition of Proof shall comply with the specification [VC_Card].¶

6.1. Self-Declaration Mode

In this mode, intelligent agents interconnect externally provided resources (including information, interfaces, etc.) using Linked-Data technologies, forming a networked ecosystem through agent description documents. Other agents can selectively retrieve appropriate resources via metadata described in these agent profile documents. Advantages of the Self-Declaration Mode:¶

• Compatibility with Existing Internet Architecture: Facilitates search engine indexing of agent-publicized information, enabling the creation of an efficient agent data network.¶

• Enhanced Privacy Protection: Pulling remote data to local systems for contextual processing mitigates user privacy leakage risks inherent in task-delegation models.¶

• Inherent Hierarchical Structure: Supports scalable interactions among a large number of agents.¶

6.2. Centralized Registration Mode

In this mode, the AI Agents register their attributes to a centralized Agent Registration Server. The parameters that an Agent needs to register in a trust domain (step A in Figure 1) may include:¶

• Name: The name of the Agent, which may not be unique and typically represented as a string.¶

• Digital Identifier: The global unique ID of the Agent configured by the network provider.¶

• Description: provide a more concise summary of the Agent’s relevant details based on natural language.¶

• Address: The access address of the Agent, which might be an URL, FQDN, etc.¶

• Version: The current version of the Agent.¶

• Capabilities: The capabilities supported by the Agent, including the communication capabilities, interaction modes and multimodal capabilities, etc. The communication capabilities refer to the communication protocols supported by the Agent, such as http/2, http/3, A2A, ANP, MCP, etc. The interaction modes may include request-response and subscription-notification and others. The multimodal capabilities refer to the data modalities that the agent can process, such as text, images, video, real-time audio, etc.¶

• Services: The services that the agent can provide. E.g., AI service, sensing service, computing service.¶

• Skills: A list of detailed description of the skills supported by the Agent. The content of each skill includes the name, ID, corresponding services, brief abstract, required input, etc.¶

• Interfaces: The APIs interfaces that the agent can provide.¶

• Security related information: For example, the licenses, authentication credentials, keys of the Agent.¶

7.1. Proactive Discovery Mode (Corresponding to Self-Declaration Mode)

In this operational mode, AI agents dynamically acquire Agent Description (AD) documents from peer agents through standardized discovery protocols (e.g., search engine). These AD documents serve as structured entry points for targeted crawling operations within Linked Data networks. The crawling mechanism implements selective resource retrieval, encompassing both semantic information and service interfaces, while adhering to ethical crawling policies.¶

7.2. Centralized Query Mode (Corresponding to Centralized Registration Mode)

In the mode, the discovery of AI agents depends on the Agent Registration Server, and the discovery process consists of two phases: "query matching" and "result feedback":¶

1) Query Matching Phase:¶

The initiating AI Agent A send a Capability Discovery request to the Registration Server, and the server screens and matches the target agents based on the capability database. The request parameters should be structured (to avoid ambiguous descriptions). Examples are as follows:¶

- Requirement description: "Medical image analysis"

- Location range: "Within 1 kilometer of base station BS-001"

- Real-time requirement: "Latency ≤ 100ms"

- Security level: "Medical qualification VC is required"

The Agent Registration Server matches the requirement with local registered Agent description. After the matching is completed, a "target agent list" is generated, which includes Digital Identifier, Address, and Capabilities, etc.¶

2) Result Feedback Phase:¶

The Agent Registration Server feeds back the matched results to the initiator AI agent, and the initiator selects a Agent based on the results and starts the session establishment with that Agent.¶

8.1. Overview

The core function of a task is to enable the AI agents involved in the communication to agree on "what to do", thereby avoiding collaboration failures due to misunderstandings. Tasks can be used in capability discovery and communication procedures:¶

• Capability Discovery: Obtaining AI agents with matching capabilities based on the task descriptions.¶

• Communication: When a Coordinator Agent initiates a communication request to an Execution Agent, the request message may carry a task description. In addition, other auxiliary information such as images, videos, files, can also be sent along with the task description to help accomplish the task.¶

An example as shown in Figure 2, a task can be executed by an AI agent (e.g., task0 sent to Agent B). When a complex task is received by an AI agent, this task can be broken down into a series of subtasks (e.g., task0 broken down to sub-task1 and sub-task2) with a clear execution sequence, known as a task chain, and executed by a group of AI agents (e.g., sub-task1 sent to Agent B, sub-task2 sent to Agent C). Task chain allows multiple AI agents to execute different tasks in a specific sequence based on policy, and enable multiple AI agents collaboratively to accomplish a complex task. The Agent communication protocol should support to encapsulate the task chain information, e.g., independent with the underlying network transport (e.g., IP, MPLS).¶

+---------+                                  +---------+
|         |------------send task0----------->|         |
| Agent A |                                  | Agent B |
|         |<-task status/result notification-|         |
+---------+                                  +---------+

+---------+                                  +---------+
|         |-----------send sub-task1-------->|  Agent  |
|         |<-task status/result notification-|    B    |
|         |                                  +---------+
| Agent A |                                  +---------+
|         |-----------send sub-task2-------->|  Agent  |
|         |<-task status/result notification-|    C    |
+---------+                                  +---------+

Tasks can be sent along with the message that establish communication session between AI agents, or separately using the established session between AI agents. In the communication session between AI agents, one or more tasks can be included, which may be independent of each other or associated through context. A task is identified by a global unique task ID. The task ID is generated by the agent that creates or assigns the task and are sent along with the task to the agent responsible for executing it.¶

8.2. Task States

Based on the length of time to complete the tasks, the task can be categorized into:¶

• Short-term tasks: These tasks can be quickly executed and completed, often used for simple tasks such as query the weather.¶

• Long-term tasks: These tasks require a longer period of time or involve multi-round interaction or extended waiting periods, such as writing an article. During the execution of long-term tasks, AI agents can synchronize task states or intermediate results among them as needed.¶

The task states are maintained by the execution AI agent, and the task status can be synchronized among Agents as needed.¶

8.3. Task coordination

The AI Agent communication protocol design MUST consider support for Agent Communication Server to facilitate task message forwarding. Agent Communication Server SHOULD prioritize message scheduling and forwarding based on task requirements to ensure efficient agent collaboration and meet transmission QoS objectives.¶

This prioritization scheme ensures that critical messages receive preferential treatment during congestion or resource contention scenarios.¶

When delegating tasks to Execution Agents, the Coordinator Agent may include task-relevant contextual about the contact information of the end user, the task itself, the historical preference information known by the Coordinator Agent, and other necessary conversation data, to facilitate the task execution. For example, in trip planning case, this may encompass historically booked flight/hotel preferences or dynamically perceived context like recent user dialog. The AI agent protocol should consequently support context sharing mechanisms through standardized definitions of context types, length constraints, and encoding formats to enhance the effectiveness of task execution.¶

9.1. Point-to-Point Communication

Direct Communication: AI agents directly send and receive protocol messages without the need for intermediate nodes for processing, or AI agents are unaware of these intermediate nodes. Indirect Communication: Communication between AI agents requires processing/relaying by the Agent Communication server, and the AI agent must be aware of and interact with the Agent Communication server. The function of the Agent Communication server includes but is not limited to:¶

• AI agent access control (allowing or blocking an AI agent's messages based on its identity or permissions).¶

• Application Layer Proxy (to facilitate monitoring/auditing of AI agent communication behavior, or to hide AI agent identity, etc.).¶

• Relay (to forward communication messages, making cross-domain communication easier, etc.).¶

• Traffic aggregation (to provide a tree-structured traffic regulation, improving communication efficiency).¶

• handle authentication and message relaying between the two communicating parties.¶

9.2. Group Communication

To better accomplish communication collaboration, agents can dynamically form groups. Information sent by an agent within a group can be received simultaneously by other agents in the same group.¶

9.3. PUB/SUB Communication

In this mode, the AI agent sending the information does not know which AI agents need to receive it. It first Publishes the information to Agent Communication Server, and this Agent Communication Server then distributes the information to the subscribing Agents based on their Subscribe status. Pub/Sub communication is a common and efficient method of information distribution, especially suitable for large-scale group communication scenarios.¶

11.1. Session Establishment and Control

Before communicating with Agent B, Agent A should first establish a secure connection with the Agent Communication Server. Prior to this, Agent A must undergo authentication by the Agent Communication Server. Similarly, Agent B also needs to be authenticated by the Agent Communication Server to establish a secure connection.¶

Therefore, the Agent Communication Server needs to support the states maintenance of the registered Agents, such as the states of Agent A and Agent B.¶

In order to communicate with Agent B, Agent A initiates a session establishment request to the Agent Communication Server. After verifying its permissions, the Agent Communication Server proceeds to establish the session, for example, by assigning a globally unique Session ID to the new session. This ID will be used throughout the entire session lifecycle to correlate all activities and data. Correspondingly, the Agent Communication Server needs to maintain a session table, which includes information about all Agents involved in the session, especially information about the session initiator.¶

11.2. Differentiated QoS Guarantees

During the session establishment, Agent A can provide the relevant QoS requirements for the session. Consequently, the Agent Communication Server can prioritize the processing and forwarding of messages according to these requirements to ensure the session's QoS.¶

12.1. Agent ID-based Route look-up

The scenario described in this section is when an Agent sends a message to another Agent (or a group of Agents), and the sending Agent knows the recipient Agent's ID or Group ID. According to the two major types of communication modes in Section 6, the situations can be classified as follows:¶

• Point-to-Point Communication (P2P):¶

• In the direct communication mode, the Agent looks up the corresponding IP address using the recipient's ID, thus allowing the message to be sent to the recipient.¶

• In the indirect communication mode, the Agent can delegate the ID lookup task to the Agent Communication Server, which is then responsible for sending the message to the recipient Agent based on the ID.¶

• Group Communication:¶

• The Agent delegates the Group ID lookup task to the Agent Communication Server, which is then responsible for sending the message to the recipient Agent in the same group.¶

12.2. Semantic-based Route resolution

The scenario described in this section is when an Agent wants to communicate with other Agents that possess a certain capability or attribute, but does not yet know their IDs. In this case, a semantic search system is needed to search for the Agent IDs that meet the criteria based on the capabilities or attributes described by the Agent. The message is then routed according to the retrieved ID.¶

18.1. Normative References

[RFC2119]

Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, <https://www.rfc-editor.org/info/rfc2119>.

[RFC8174]

Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017, <https://www.rfc-editor.org/info/rfc8174>.

18.2. Informative References

[I-D.stephan-ai-agent-6g]

Stephan, E., Schott, R., Lopez, D., Duan, X., and L. Morand, "AI Agent protocols for 6G systems", Work in Progress, Internet-Draft, draft-stephan-ai-agent-6g-00, 7 July 2025, <https://datatracker.ietf.org/doc/html/draft-stephan-ai-agent-6g-00>.

[VC_Card]

"Verifiable Credential Data Integrity 1.0", 15 May 2025, <https://www.w3.org/TR/vc-data-integrity/#defn-domain>.