OAuth2.0 Extension for Multi-AI Agent Collaboration

AI agents are capable of handling tasks. They can integrate multiple technologies such as natural language understanding, data analysis, and logical reasoning to meet multi-step and cross-scenario requirements. However, a single AI agent has a limited knowledge scope and restricted functions, making it difficult to independently complete complex tasks that span multiple professional domains and require different capabilities. Collaboration among a group of AI agents is an essential approach for such complex tasks. Taking the demand for "real-time health advice" as an example, this intent can be split into three tasks: "collect the user's health data", "predict the user's health status" and "provide health advice based on health status". These tasks need to be accomplished by corresponding professional sub-AI agents: the data collection agent is responsible for privacy-compliant data collection, the health status prediction agent invokes medical models for analysis, and the advice generation agent outputs solutions based on health guidelines. What's more, according to some existing research, a leading agent is necessary for coordination. In the above example, the leading agent can receive the intent "give me real-time health advice" from user, understand it and split it into tasks. Additionally, the leading agent selects sub-agents capable of fulfilling the tasks and distributes these tasks to them. These sub-agents can be referred to as a task group. Finally, the leading agent integrates the results returned by the sub-AI agents to form complete health advice. Without the leading agent, task execution is prone to confusion and final results are difficult to unify. There are two typical policies for the leading agent to select sub-agents and form a task group: 1) The static policy determines all tasks and sub-agents upfront before execution, enabling predictable scheduling and stable group composition; 2) The dynamic policy incrementally identifies tasks and selects sub-agents during execution, allowing real-time adjustment based on intermediate results and improving adaptability to changing conditions. In such collaborative scenarios, if each sub-AI agent in a task group applies for permissions from the authorization server individually, the authorization server must verify the identity and permissions of each sub-agent one by one and generate access tokens for them. This leads to problems such as frequent interactions, high computing resource consumption on the authorization server, and low efficiency in the token issuance process. In addition, when each sub-agent is authorized independently, there is no single point that bounds the total authority held by the task group, so the authority collectively exercised by the group can exceed what was intended for the task as a whole. Therefore, this draft proposes authorization methods for static and dynamic task groups. For the static policy, the leading agent communicates with the authorization server centrally and applies for access tokens on behalf of all sub-agents in the task group before task execution begins. For the dynamic policy, the leading agent applies for an access token for a specific task and issues task credentials bound to the access token to the corresponding sub-agents. In addition, this draft describes an optional scope-bounded mode, where the authorization server issues a group authorization token and member authorization tokens; the authority of each member, and the aggregate authority of all members, is bounded by the authority granted to the group. These methods simplify the authorization process for sub-AI agents and avoid efficiency loss caused by repeated interactions.

Applier: a role that requests the access token on behalf of client(s), e.g. a leading AI agent. Group Authorization Token: a token issued by the authorization server that authorizes the group permission scope, i.e. the maximum authority the whole task group may exercise while handling a task. Member Authorization Token: a token issued by the authorization server that authorizes the member permission scope of an individual sub-agent. The member permission scope is no broader than the group permission scope. Group Permission Scope: the set of resources, the set of service types, the set of permitted operations, and the maximum access count that the task group as a whole is permitted while handling a task. Member Permission Scope: the resources, service types, operations, and maximum access count permitted to a single sub-agent. It is no broader than the group permission scope. This draft uses the terms "authorization server", "client", "resource server" defined by OAuth 2.0 .

This protocol flow is modeled after and extends it with an *applier_id* field to enable authorization request by the applier.

| Resource | | | | Owner | | |<--(B) Authorization Grant----| | | | +----------------+ |Leading | | Agent | (C) Authorization Grant +----------------+ | |--------(applier_id)--------->| | | | | Authorization | | | (D) Access Token | Server | | |<------(app=applier_id)-------| | +--------+ +----------------+ | | (E) Access Token* | v +--------+ +----------------+ | |-------(F) Access Token*----->| Resource | | Sub- | | Server | | Agent |<---(G) Protected Resource----| | +--------+ +----------------+]]>

Before initiating the protocol, the leading agent registers with the authorization server, as defined in . In this draft, the leading agent's role or capability is incorporated into its registration process. For example, the leading agent's profile includes a "capability" parameter with the value "resolve intent and distribute tasks".

This section refers to step (A) and (B). The leading agent receives a requirement from the user, resolves it and determines the task(s) and sub-agent(s) needed to fulfill the requirement. For example, when the user sends the intent "give me real-time health advice" to the leading agent, the leading agent resolves this intent and identifies three tasks: task 1 (collect the user's health data), task 2 (predict the user's health status) and task 3 (provide health advice based on health status). The leading agent then utilizes the agent discovery process to select sub-agent(s) capable of performing the respective tasks. For instance, task 1 assigned to sub-agent 1, task 2 to sub-agent 2, and task 3 to sub-agent 3. In the subsequent authorization process proposed in this draft, the leading agent is designated as the applier, and the selected sub-agent(s) are designated as the client(s). The leading agent requests authorization from the resource owner, as defined in .

This section refers to step (C) and (D). The leading agent requests an access token by authenticating with the authorization server and presenting the authorization grant, as defined in . The message may also include the following parameter: OPTIONAL. The identifier of the applier which is requesting access token on behalf of the client(s). The authorization server authenticates the leading agent and validates that the leading agent is capable of requesting access tokens on behalf of sub-agent(s). For example, it checks whether the leading agent's profile includes a "capability" parameter whose value includes "distribute tasks". If the validation is successful, the authorization server may verify grants as defined in . Then the authorization server generates access token and send access token response message to the leading agent. The access token is similar to OAuth2.0, except that it consists of an additional claim to specify the applier. OPTIONAL. The identifier of the applier which is requesting access token on behalf of the client(s). The access token includes one *app* parameter and multiple *sbj*-*aud*-*scope* grants. Below is an example: { "iss": "authorization server ID", "app": "leading agent ID", "grants": [ { "sbj": "sub-agent 1 ID", "aud": ["resource server 1 ID"], "scope": "..." }, { "sbj": "sub-agent 2 ID", "aud": ["resource server 1 ID","resource server 2 ID"], "scope": "..." }, { "sbj": "sub-agent 3 ID", "aud": ["resource server 2 ID","resource server 3 ID"], "scope": "..." } ] } If the validation fails, the access token response message may include the reason for failure. has defined types of error response. In this case, the authorization server may use a new error message "unauthorized_applier" to indicate that the leading agent is not capable of requesting access tokens on behalf of sub-agent(s).

This section refers to step (E). The leading agent sends the task ID and access token* to the sub-agent(s). Based on local policies and regulations, the leading agent may use privacy protection mechanisms to process the access token. Thus, the access token* may be the same as the access token, or generated from the access token using privacy protection mechanisms. For example, the leading agent uses a selective disclosure algorithm to generate access token 1, access token 2, and access token 3 from the access token received in step (D). Access token 1 may include the leading agent ID and sub-agent 1 ID; access token 2 may include the leading agent ID and sub-agent 2 ID; access token 3 may include the leading agent ID and sub-agent 3 ID. The applier may then send access token 1 to sub-agent 1, access token 2 to sub-agent 2, and access token 3 to sub-agent 3. The sub-agent(s) validate the proof of the received access token* and check whether the applier ID refers to a trustworthy leading agent. For example, each AI agent is preconfigured with a list of trusted AI agents, and any one of the trusted AI agents that acts as a leading agent will pass the validation. If the validation fails, the client(s) may send a response message to the applier with a failure indication of "unknown_applier".

This section refers to step (F) and (G). The sub-agent(s) execute the task according to message received in step (E). When resources are needed, the sub-agent(s) request the protected resource from the resource server and present the access token*, as defined in . The resource server validates the access token*, as defined in . If the access token* includes multiple *sbj*-*aud* pairs, the resource server may verify that its own ID and the sub-agent's ID are in the same pair. In the previous example of the access token, "sub-agent 1 ID" and "resource server 1 ID" are in the same pair. If the validation succeeds, the resource server may provide the resource to the client(s).

This protocol flow is modeled after and extends it with task credentials bound with the access token.

| Resource | | | (B) Authorization Grant | Owner | | |<-----------------------------| | |Leading| +-------------+ | Agent | (C) Authorization Grant +-------------+ | |--------(task, key)---------->| | | | |Authorization| | | (D) Access Token | Server | | |<-------(task, key)-----------| | +-------+ +-------------+ | (E) Access Token* & Task Credential | +---v---+ +-------------+ | | (F) Access Token* | | | Sub- |----------------------------->| Resource | | Agent | (G) Protected Resource | Server | | |<-----------------------------| | +-------+ +-------------+]]>

Before initiating the protocol, the leading agent registers with the authorization server.

This section refers to step (A) and (B). The leading agent receives a requirement from the user, resolves it and determines the task(s) and client(s) needed to fulfill the requirement. The leading agent requests authorization from the resource owner. Detailed procedures are similar to Section 3.2.

This section refers to step (C) and (D). The leading agent requests an access token by authenticating with the authorization server and presenting the authorization grant, as defined in . The message may also include the following parameter: OPTIONAL. The identifier or description of the task which is resolved from intent. OPTIONAL. The index that point to the public key of the leading agent that is used to issue task credentials. The authorization server authenticates the leading agent and validates that the leading agent is capable of executing specific task. What's more, the authorization server validates that the leading agent holds the private key related to the *key* received in the request message. If the validation is successful, the authorization server may verify grants as defined in . Then the authorization server generates access token and send access token response message to the applier. The access token is similar to OAuth2.0 and OAuth2.0 DPoP, reuse the *subject* parameter to describe the access token is issued to agents that can execute specific task, and add a new parameter to point to the task credential (e.g. the public key used to issue task credentials). OPTIONAL. The indication of public key of the leading agent that is used to issue task credentials. Any agent that holds the credential issued by this public key is authorized to request resources defined in the access token. Below is an example: { "iss": authorization server ID, "sbj": task ID, "aud": resource server 1 ID, "key": public key for specific task ... } If the validation fails, the access token response message may include the reason for failure. In this case, the authorization server may use a new error message "unrecognized_pk" to indicate that the validation of *key* sent by the leading agent is wrong.

This section refers to step (E). After the leading agent selects sub-agent to execute task, it generates task credential and sends the credential with access token* to the sub-agent(s). The task credential is used to prove that the subject of the credential is selected by the leading agent to execute specific task. Optionally it includes the information of the related access token. The credential is signed by the leading agent using the private key related to the public key indicated by the *key* parameter in the access token. This credential can have multiple formats. Below is an example: { "issuer": leading agent ID, "subject": sub-agent ID, "claim": task ID "related token": hash of the access token ... "proof": signature of leading agent } The leading agent may resolve a list of multiple tasks based on one intent from the user. In this case, the *sbj* in the access token may involve several tasks. Based on local policies and regulations, the leading agent may use privacy protection mechanisms to process the access token. For example, the leading agent uses a selective disclosure algorithm to generate access token 1, access token 2, and access token 3 from the access token received in step (D). Access token 1 may include task ID 1; access token 2 may include task ID 2; access token 3 may include task ID 3. The leading agent may then send access token 1 to sub-agent 1, access token 2 to sub-agent 2, and access token 3 to sub-agent 3. The sub-agent(s) validate the received access token* and task credential, and validate the relationship of them. For example, to validate the task credential, the sub-agent(s) check whether the *subject* indicates a trusted leading agent, and whether the *proof* value is right; to validate the relationship, the sub-agent(s) check whether the task ID in the access token* is the same as the task credential, whether the key indicated by the access token* can be used to validate the proof of the task credential, and whether the token hash in the task credential is same as the hash of the access token*. If the task credential validation fails, the sub-agent(s) may send a response message to the applier with a failure indication of "invalid_credential". If the relationship validation fails, the sub-agent(s) may send a response message to the applier with a failure indication of "unknown_credential".

This section refers to step (F) and (G). The sub-agent(s) execute the task according to message received in step (E). When resources are needed, the sub-agent(s) request the protected resource from the resource server and present the access token*, as defined in . Meanwhile, the sub-agent(s) send the task credential to the resource server. The resource server validates the access token*, as defined in , and validates the task credential and the relationship between the access token* and the task credential. If the validation succeeds, the resource server may provide the resource to the client(s).

This section describes an optional enhancement to the static and dynamic task group authorization in Section 3 and Section 4. It is not a third task group formation policy; rather, it strengthens either policy by bounding the authority delegated to the task group as a whole and to each member. The authorization server issues a group authorization token that caps the authority of the whole group, and a member authorization token for each sub-agent whose authority is no broader than, and in aggregate does not exceed, the group's authority. This realizes least privilege at both the group and the member level and limits the impact if a single sub-agent is compromised.

| Resource | | |<-(B) Authorization Grant------| Owner | | | +----------------+ | | |Leading | (C) Token Request +----------------+ | Agent |--(group req + member req(s))-->| | | | | Authorization | | | (D) Group Token + | Server | | |<----Member Token(s)-----------| | | | | | | | (H) Revocation Request | | | |--(group / member token)------>| | +--------+ +----------------+ | (E) Member Token (+ group token info) | v +--------+ +----------------+ | Sub- |--(F) Member Token------------>| Resource | | Agent |<-(G) Protected Resource-------| Server | +--------+ +----------------+]]> Step (H) is performed when a sub-task or the whole task is completed, or when policy requires token revocation; it is not part of the initial token issuance.

This section refers to steps (A) through (E). Before requesting tokens, the leading agent registers with the authorization server as in Section 3.1. The leading agent's profile may indicate whether it is authorized to manage a task group (e.g. a "capability" value "manage task group"). The leading agent resolves the user's requirement into one or more sub-tasks and selects the sub-agent(s) for them. The selection of sub-agents for sub-tasks is out of scope of this document; the leading agent may determine the sub-agents locally or use an external discovery or management function. The leading agent then obtains an authorization grant from the resource owner as in Section 3.2. When scope-bounded task group authorization is used, the token request to the authorization server includes the following parameters: Information used to request the group authorization token. It includes the task information, and may include the group information and the leading agent information. One or more entries, each used to request a member authorization token. Each entry includes the sub-agent information, and may include the group information. The authorization server authenticates the leading agent and validates that it is capable of requesting tokens on behalf of a task group, as in Section 3.3. If validation succeeds, the authorization server generates a group authorization token based on the task information, and a member authorization token for each sub-agent based on the sub-agent information and the group permission scope. The authorization server ensures that each member permission scope is within the group permission scope: the member resource set, service-type set, and operation set are each no broader than the corresponding group set. For quantitative limits, such as access count or quota, the sum of the limits assigned to all member tokens does not exceed the corresponding limit of the group token. The group permission scope and member permission scope are each expressed in terms of one or more of: the set of resources, the set of service types, the set of permitted operations (e.g. read, query, update, subscribe), and a maximum access count. A minimal example, in which the member token is bounded by the group token, is: group token: "grp":"G1", "scope": { "resources":["r1","r2"], "operations":["read","update"], "max_calls":100 } member token: "grp":"G1", "sbj":"A1", "scope": { "resources":["r1"], "operations":["read"], "max_calls":20 } The authorization server returns the group authorization token and the member authorization token(s) to the leading agent. If a requested member permission scope cannot be bounded by the group permission scope, the authorization server may return a new error "scope_exceeds_group". The leading agent then delivers each member authorization token to the corresponding sub-agent, together with the sub-task identifier. The leading agent may also provide a reference to the group token when the sub-agent needs to verify that the member token is associated with a valid task group.

This section refers to steps (F) and (G). The sub-agent executes its sub-task. When resources are needed, the sub-agent requests the protected resource from the resource server and presents its member authorization token, as defined in . The resource server validates the member authorization token and verifies that the requested resource, service type, and operation are within the member permission scope, and that the member's access count has not been exceeded. The resource server may additionally verify, from the "grp" claim, that the member belongs to the expected task group. If the requested access is outside the member permission scope, the resource server denies the request.

This section refers to step (H). When the whole task is completed, the leading agent sends a revocation request to the authorization server to revoke the group authorization token, optionally together with the member authorization token(s). When a sub-task is completed, the leading agent sends a revocation request to revoke the member authorization token of the sub-agent that executed the sub-task. Upon receiving a revocation request, the authorization server verifies that the leading agent has the permission to manage the task group (see Section 5.1) before revoking the token(s). After revoking a member authorization token, the authorization server may release the corresponding sub-agent instance. The group authorization token and member authorization token may be configured with a validity period, based on which the authorization server maintains token state. The leading agent may also send sub-task execution information that the authorization server uses to maintain the member authorization token(s).

This section discusses security considerations specific to multi-agent task group authorization.

The two-tier token model in Section 5 enforces least privilege at two levels. The group authorization token bounds the total authority delegated to a task group, and each member authorization token is no broader than the group permission scope. Because the authorization server ensures that the member scopes are within the group scope and that quantitative limits sum to no more than the group's, a leading agent cannot, by partitioning work across many sub-agents, cause the group to collectively exercise more authority than was granted to the group as a whole.

A compromised sub-agent can only act within its member permission scope, which is typically much narrower than the group permission scope. This limits the impact of a single compromised member. Per-sub-task member token revocation (Section 5.3) further reduces exposure by removing a member's authority as soon as its sub-task completes, rather than waiting for the whole task to finish.