Build a Course Enrollment Agent

Build an agent that helps students find and enroll in courses. Ask “List the available courses”, then “Enroll me in CS205.” The first request uses the agent's own identity. The second pauses for your sign-in and consent before the agent enrolls you.

Set up the identities, permissions, and consent flow once, then choose LangChain, Google ADK, Vercel AI SDK, or CrewAI. Each quickstart implements the same conversation with two tools: list_courses and enroll_course.

Understand the Two Flows

An agent identity identifies the agent itself in ThunderID. Registering an agent gives it an Agent ID, plus a Client ID and Client Secret it uses to request access tokens. An access token tells the tool whose authority it is using and which permissions it has.

On behalf of a user (OBO) means the agent acts with a user's delegated authority. Signing in identifies the student; consent gives the agent permission to act for that student. The resulting token names both parties.

Request	Whose authority?	Permission	Student involvement
“List the available courses.”	The agent's own identity	courses:read	No sign-in or consent needed.
“Enroll me in CS205.”	The signed-in student's delegated authority	courses:enroll	Sign in and approve the requested permission.

The agent selects a tool from the request. The runtime obtains the token required for that operation without exposing it to the model. The tool validates the token and checks its permission before performing the action, so the model cannot select a student ID or supply an access token.

The agent runtime and both course tools run in the same sample application process.

List Courses as the Agent

The agent requests a token with its Client ID and Client Secret using the client credentials grant. The token represents the agent and carries courses:read. The listing tool validates it with ThunderID, checks that permission, and returns the catalog.

A student can ask the question without delegating their identity. It is the agent's role that grants access to the catalog.

Enroll the Student with OBO

When you ask “Enroll me in CS205”, the agent selects enroll_course. Its runtime wrapper requires delegated authority, so it starts an authorization code flow with PKCE before invoking the tool. You follow the sign-in link to ThunderID, sign in, and approve courses:enroll on its consent screen.

ThunderID returns an authorization code through your browser to the agent's registered redirect URI. The runtime exchanges it for a delegated access token and passes that token to the tool. PKCE binds the exchange to the request that started it. The tool validates the token and its permission, then enrolls the student identified by the token.

If you deny consent, the configured flow fails and the agent does not enroll you. Closing the browser leaves the runtime waiting until its 180-second timeout. A token without courses:enroll also fails the tool's permission check.

Read the Identity in the Result

The sample prints the following fields after token validation:

Field	Listing courses	Enrolling a student
sub	Agent ID	Signed-in student ID
act.sub	Absent	Agent ID
client_id	Agent's Client ID	Same agent's Client ID
scope	courses:read	Includes courses:enroll

The subject (sub) is whose authority the token represents. The actor (act.sub) is the agent exercising delegated authority. In OBO, the enrollment belongs to the student, while the actor identifies which agent performed it.

Authentication and authorization serve different purposes here. Signing in proves who the student is. Their role determines whether they may enroll, consent approves the agent's use of that permission, and the tool checks the granted permission before changing the record. Consent cannot grant a permission the student does not hold.

Register the Agent

Start ThunderID if you do not have an instance running, then open its Console at https://localhost:8090/console.

$npx thunderid

Requires Node.js 18+

Full install guide →

1. In the Console, open Agents and click Add Agent.
2. Enter Course Enrollment Agent. Select an organization unit if prompted, then click Continue.
3. For Model Provider, select Gemini. Enter gemini-flash-lite-latest for Model, then click Continue. These fields describe the agent; your framework uses the API key you configure later.
4. Select an Owner, then click Create agent.
5. Copy the secret from Save your client secret, then click Continue. The secret appears only once.
6. On Overview, copy the Client ID from Agent details. Keep it and the secret for your framework's .env file.
7. On Advanced, confirm Client Authentication Method is client_secret_basic. Turn on Delegated mode, keep Require PKCE on, and add http://localhost:6274/callback under Authorized redirect URIs. Leave Require Pushed Authorization Requests off for this sample and save your changes.

The same agent supports both client_credentials and authorization_code. The browser and sample must run on the same computer so the redirect can reach the local listener.

Define the Course Permissions

A resource server groups permissions for a protected service. Here it represents the sample's course tools, even though those tools run locally.

1. Open Resource Servers and create a resource server of type API. Name it Course Catalog, set Identifier to https://courses.example.com, and use : as the permission delimiter. Keep the existing default resource server unchanged.
2. On its Resources tab, create a resource named Courses with handle courses.
3. Under that resource, create the following actions:

Action name	Handle	Generated permission
List courses	read	courses:read
Enroll in a course	enroll	courses:enroll

https://courses.example.com is an identifier, not a service you need to deploy or visit. The samples send it as resource when requesting tokens and check it as the token's audience. Set COURSE_RESOURCE to this same value in each sample.

See Manage Resource Servers for the API equivalent.

Assign the Agent and Student Roles

Use an existing non-admin user as the student, or create a test user with a username and password. Course enrollment does not create a user account.

1. In Roles, create Course Reader in your organization unit. Select courses:read from Course Catalog as its permission.
2. On that role's Assignments tab, click Add Assignment. Select Agent and assign Course Enrollment Agent.
3. Create another role named Student with courses:enroll from Course Catalog.
4. On Student → Assignments, click Add Assignment. Select User and assign your test student.

The agent can read courses under its own identity. Enrollment requires the student's role and consent; the agent's Course Reader role does not include enrollment permission.

Configure Sign-in and Consent

Create a separate authentication flow for this sample so other applications keep their existing sign-in behavior.

1. Open Flows and create a sign-in flow from the Basic template. Name it Course Enrollment Sign-in.
2. In the flow builder, add User Consent from Widgets. It creates the consent screen and its Allow and Deny responses.
3. Check the successful path: Identifier + Password → Authorization → User Consent → Auth Assertion Generator. The authorization step resolves the student's permissions before consent. Both responses from the consent screen must return to User Consent.
4. Select the User Consent executor and turn on Fail flow when user denies consent. Save the flow.
5. Return to Agents → Course Enrollment Agent → Flows. Select Course Enrollment Sign-in as the Sign-in Flow and save.

The enrollment request includes scope=openid courses:enroll, the course resource identifier, and prompt=consent. The configured flow authenticates the student, checks their permission, and asks them to approve it. prompt=consent requests a fresh consent decision; it does not create a consent step in a flow that lacks one.

The consent grants the courses:enroll permission, not a course-specific restriction to CS205. The sample obtains a new delegated token for each enrollment request and keeps it only for that tool call. A production agent may reuse an unexpired token within the user's session according to its consent policy.

See Configure consent for the flow builder details and Agent access on behalf of users for other delegation methods.

Next Steps

Choose your framework to create the tools, connect them to an agent, and try both requests:

LangChainBuild the agent with LangChain in Python or TypeScript.Open LangChain guide →

Google ADKBuild the agent with Google ADK in Python.Open Google ADK guide →

Vercel AI SDKBuild the agent with Vercel AI SDK in TypeScript.Open Vercel AI SDK guide →

CrewAIBuild the agent with CrewAI in Python.Open CrewAI guide →

Each example uses real ThunderID tokens, consent, and permission checks. Course records live in memory and disappear when the process exits. When you move the tools behind an API, validate tokens and enforce permissions in that API; see Control agent access.