Multiple Agents Collaboration in 6G Network

3GPP research follows sequential phases spanning use cases, requirements, key issues, solutions, and protocol implementation. Different Working groups are responsible for different phases. In order to accelerate the commercial deployment, the timeframes for each phase may partially overlap. For example work on architectural requirements and key issues identification kicks off as the use cases and requirements approaching consolidation. 3GPP SA1 started the study on 6G use cases and service requirements in Q4 of 2024 and completed the evaluation and conclusions in Q1 of 2026. Consolidated potential functional and performance requirements for AI are achieved in clause 14.2 in . Standardization work for 6G Service Requirements started in Q2 of 2026 and were addressed in (considered as a working draft and non-published specification). Six functional requirements have been listed in clause 5.2 in and update of requirements are still under discussion. 3GPP SA2 started the study on Architecture for 6G System in Q3 of 2025. Two main key issues, i.e. AI for 6G architecture and 6G Network for AI, are decomposed into several sub-KIs in clause 5.18 and 5.19 of . Because there are so many solutions for these KIs, clause 6.18 and 6.19 only show the summary of solution variants and the detailed solutions are listed in Annex C.18 and C.19. Details of key issues and solution variants are described in section 3 and section 4. 3GPP CT3 started the study on the Protocol for AI in the 6G System in Q2 of 2026. It focuses on protocol design considerations and candidate protocols before SA2 reaching any consolidated agreement. Details of protocol design considerations and candidate protocols are described in section 5. progress in other WGs of 3GPP will be added later.

This section describe considerations for Agent communication involving 6G network. Figure 1 shows agent communication among the agents on the UEs, in the mobile operator network (6G network) and on the internet.

The functionalities described in Figure are as follows: UE side Agent Application can be installed in the UE and communicate with agent or invoke tool/skill in the network. Communication Modem is used to communication with the mobile operator network support the functionalities such as baseband processing, NAS message transfer, and etc. In 6G, the request with intent may be handled in communication modem. Mobile operator network: The NFs (Network Function) exchange singalling with each other via the service based interfaces (SBI based on HTTP2). These NFs and interfaces between them consist of the control plane. The NFs transfer the data for UE via the user plane. Multi-agent is merely a schematic, several variants for collaboration will be discussed in section 3. Service: Comparing with agent application, a 3rd party AF is the entity that has been authenticated and authorized by the network operator to access specific network services or resources. Depending on where agents reside, the agent communication involving 6G network can be classified as follows: UE -- Service: UE may install the Agent Apps and communicate with the Agents/Tools on the internet/cloud. Data stream of such communication may be transferred via the mobile operator network(6G network). The communication protocol can re-use the existing open-resource protocols. However, in order to ensure the user experience, the 6G network collects information of various traffic characteristics (e.g. modality, tolerated error rate, priority, generation/arrival rate of data) of such communication. Based on collected information 6G network may: adjust the transmission policy. predict communication patterns. expose the traffic characteristics (e.g. bursty data characteristics) to authorized 3rd party to optimize communication services. UE -- Mobile Operator Network: UE communicates with the Agent/tool in the mobile operator network. When providing AI service to the UE, the 6G network need to consider: whether and how to allocate the identity for Agent communication. providing 3GPP agent service(s) to the authorized UE. authorizing the agents residing on the UE and 3rd party. exchanging message to the UE with or without intent. AI model training for UE, e.g. collect inform from the UE and download the trained model to UE. Intra Mobile Operator Network: UE communicates with the Agent/tool in the mobile operator network. When providing AI service to the UE, the 6G network need to consider: Agent/tool discovery within 6G network. Agent/tool registration may be needed before it can be discovered. intent translation from the UE or authorized 3rd party to one or more 3GPP services. Network resilience, e.g. network failure detection, recovery, load balance and etc. AI model training and inference. 3rd party AF -- Mobile Operator Network: Authorized 3rd party AF/agent/tool is allowed to communicate with 6G network and 6G network may: expose 3GPP agent service/tool to the authorized 3rd party AF. invoke AI capabilities of the authorized 3rd party AF (i.e. tool or agent) if 6G network cannot fulfill the UE request. exchange message to the 3rd party AF with or without intent. Based on the functionalities above, it is evident that 6G networks play an integral role in the agent ecosystem. Agents residing in 6G network not only provide AI capabilities (agents or tools) to the UE and the 3rd party but also play a key role to ensure efficient and secure transport for agent communication. Furthermore, the mobile operator network has to consider stability, robustness, compatibility and evolutionary. Thus the existing open-resource protocols will not be considered as candidate protocols for agents directly. In order to integrated into the agent ecosystem, 6G network prefers re-using IETF protocols. On the other hand, the mobile operator network with definite boundaries may also define or enhance protocols for the cases of agent communication between the UE and the mobile operator network and agent communication intra mobile operator network.

Figure 2 describes the solution variant in subclause 6.18.1.1.1 of

This solution variant aims to propose the high-level architecture to enable 6G CN to handle and fulfil the requests with intent. It enables the 6G CN to act as an operator-controlled multi-agent system to handle and fulfil the UE/AF requests with intent and consist of three parts: Agent functions: A central agent as the planner/coordinator interpret and decompose the intent, generate a plan, and monitor the plan during the lifecycle. It coordinates the capabilities provided by other agent, non-agentic NFs and trusted 3rd party AF and requests the executors to execute task based the plan. Supportive intermediary functions: three categories functions supportive intermediaries have been defined as: Dedicated entity for bridging acts as a controlled interaction point between agents and other traditional NFs executing the task allocated by planner. It supports exposing the network capabilities of traditional NF to agent, translating the invocation request from the agent to traditional NF and authorizing and validating the requests initiated by agent. Dedicated entity for intent-based access acts as a controlled access point for handling the requests including intent from UE and/or 3rd party AF. It supports validating and authorizing the request from the UE/AF, maintaining session association and routing messages between UE/AF and agents. Considering functionalities of ID allocation for the UE AI agent, authentication and authorization for the agent residing on the UE/ 3rd party AF, the access entity may be enhanced or separated entities will be added. Supporting of these functionalities will be updated based on 3GPP progress. Dedicated entity for operator control and governance acts as a deterministic enforcement node to enhance the operator governance and control of agentic capabilities. It supports auditing agent-generated intent/plans, translating AI decisions to deterministic enforcement (e.g. standardized NF service request) and monitoring of NF or network status. Repository entities enable the registration and discovery of agents or tools. The agents may be totally new functionalities. The supportive intermediary functions may co-located with agents or traditional (enhanced) 6G NFs. The repository entities may also co-located in NRF which is responsible for NF registration and discovery in 5G. In order to fulfill the intent from the UE and 3rd party NF, there are three sub-variants for invoking the traditional 6G NFs as described in subclause 6.18.16 of : Invoking defined 3GPP procedure and some NF services. It will re-use the procedure and services defined for a general control plane and doesn't depend on IETF protocols. Invoking dynamically composing procedures with tools invocations, NF services invocations and collaborations among different NFs/AFs via tasks. It will depends on agent communication defined in IETF. Invoking defined 3GPP procedures and some NF services represented as tools. Notice that the definition and granularity of tools and how to map the tools and traditional 3GPP procedure and service are also under discussion of functional solution variants. Thus the dependency of the 3rd sub-variant will be updated.

The interaction as follows is considered to use agent-based interface based on the Agent communication protocols defined in IETF: the interaction between the agent functions, i.e. planner and executor the interaction between the agent functions and 3rd party AF the interaction between supportive intermediary functions and agents The interaction as follows may use the Agent communication protocols or intent-based protocols specified in 3GPP: the interaction between the agent functions and the UE the interaction between the agent functions and the enhanced NF with tools/skills based on different sub-variants listed in section 3.1.1 the discovery of agents across repository entity boundaries, e.g. discovery of agent/tool in 6G by 3rd party, roaming cases and etc.

Figure 3 describes the solution variant in subclause 6.18.3.1.1 of

This solution variant aims to propose a new separated domain called AI domain overlaying the traditional PS domain. AI domain is defined as an operator-managed trusted domain. It enables intent-based interactions among UE AI domain client, 6G CN AI domain and AI Agents on third-party AFs. It may consist of the following functionalities: AI domain entry point: It is the first-entry point to the 6G CN AI domain for UE AI domain clients and authorised 3rd party AFs. NFs in AI domain: Some typical NFs have been listed as follows: NFs responsible for processing and fulfilling requests/intents received from UE and 3rd party. NFs as translators responsible for consuming the traditional NF capabilities in PS domain i.e. mapping the requests between the AI domain and PS domain and invoke the service exposed by the NFs in PS domain. Supporting functionalities: security, policy enforcement, context management are considered as supporting capability in this architecture. The security functionalities may include authentication and authorization for the agent residing on the UE/ 3rd party AF. More details will be updated based on 3GPP progress. re functions in AI domain will be added later based on the progress.

The interaction as follows is considered to use agent-based interface based on the Agent communication protocols defined in IETF: the interaction among the entry point, NFs and Supporting functionalities in the AI domain the interaction between the UE/3rd party AF and the entry point in the AI domain The interaction as follows is considered to use protocols specified for 6G in 3GPP. the interaction between the translator and traditional NFs in PS domain

Figure 4 describes registration, discovery and exposure of AI capibility as described in subclause 6.18.17, subclause 6.19.2, subclause 6.19.3 and subclause 6.19.4 of

There may be a common repository function or more than one repository functions in 6G core-network. Some sub-variants also depend on the architecture variants. Thus Figure 4 describes the functionalities from a general perspective. Three typical AI capabilities of NF are considered to register in one general repository or multiple dedicated repositories as describes in subclause 6.18.17.1.1 of : basic AI/ML capability: They are provided by 6G NFs (e.g. AI/ML training, AI/ML inference). They are consumed by other 6G NF (e.g. basic NF, Agentic NF, NF in AI Domain) to perform AI/ML operations. tools: They represent specific functional capabilities or standardized operations exposed by NF(s) for execution by 6G NF (e.g. Agentic NF, NF in AI Domain). The registered information includes the tool ID, tool name, tool type, tool description, required input, optional input, required output, optional output.. skills: They represent specific functional capabilities exposed by an agent. One agent may discover another agent for intent fulfilment. UE agent registration is still under discussion. It will be updated based on the latest progress as described in subclause 6.19.1.1.2 of . The NF with AI capabilities may expose the network capabilities to the AF as described in subclause 6.19.3 of and expose AI service to UE and 3rd party application as described in subclause 6.19.3 of . AI agent on AF can discovery the network capabilities via the intermediary node or to the repository function directly. Further in the sub-variant described in subclause of 6.19.3.4 of , it requires the AF AI Agent to register in the repository in 6G AI domain as described in 3.2. The well-trained model can be stored in an AI repository function and be discovered by/exposed to the agent on the UE or AF as described in subclause 6.19.4 of [TR 23.801-01v070]. How to retrieve well-trained model depends on architecture variants.

The interaction as follows is considered to use the discovery protocols defined in IETF: registration and discovery of AF AI Agent in the repository in 6G AI domain. registration and discovery of UE AI Agent in the repository in 6G AI domain. The interaction as follows may use the Agent communication protocols defined in IETF or intent-based protocols specified in 3GPP: registration and discovery of NF with AI capabilities exposure of NF AI capabilities to the AF exposure of AI service to the agent on the UE or AF

Solution variants of UE AI agent ID assignment/generation are described in subclause 6.19.1.1.1 of . In 6G network, the ID allocated/configured by the mobile operator is used to identify and authenticate the UE. It is assumed that the UE AI agent has a permanent identifier. The permanent identifier may be preconfigured in the UE AI agent or dynamically assigned by network after authentication with the external server. The permanent identifier may be mapped with an external application identifier. In this case the authentication may also involve the external server.

The interaction involving authentication by external server using an external application identifier may depends on progress, if the external application identifier allocation is specified in IETF.

Only two protocol design consideration have been addressed in clause 5.1 of . More consideration will be updated according to progress in 3GPP.

TBD