AI Agent Example

This example can be viewed from GenAI Model screen in Data Science app

This training agent is a “chat + tools” example. It searches books by calling the /train_rest saga. It can also render results as a rich HTML card UI.

It is configured in the Data Science → GenAI Models screen. It depends on the training assets from the other examples.

Before you start

• The training runners are installed and running.

• The /train_rest saga exists and is ACTIVE. See API Flow Examples.

• A runner is allowed to serve GenAI models (train_rpc). See Microservice Examples.

• You have an LLM provider key stored as a secret.

For OpenAI, store rierino.system.openai.apikey key as a secret. In local setups this is typically globalsecrets.properties. In Kubernetes this is typically the global-secrets Secret.

What you’ll build

• A GenAI agent with sample system instructions and a book-search goal.

• A tool saga entry that lets the agent call /train_rest.

• An optional tool state entry (the training dummy model) for safe additional data access.

• An optional interface so the agent can respond with UI cards.

End-to-end flow (what happens at runtime)

1. The user asks for books (author/title/keywords).

2. The LLM decides to call the /train_rest tool saga.

3. /train_rest calls Open Library via an outbound REST step.

4. The agent formats results.

5. It returns an HTML-based result card.

Configure the model (walkthrough)

1

1) Define the agent (purpose + allowed runner)

Fill the model definition. Keep the instructions concrete. Limit Allowed For to the runner you want serving the agent, unless it should be running within all microservices.

Training Agent Example

2

2) Select the LLM provider + API key secret

Pick OpenAI (or your provider). Reference the API key through a secret entry.

Training Agent LLM Model

3

3) Add tools (make it actionable)

Add the /train_rest saga as a Tool Saga. This is what turns a chat-only agent into an agent that can act.

Add dummy as a Tool State. Keep it read-only unless you need writes.

Training Agent Tools

Tool sagas work with an input schema. Schema nudges the LLM into correct parameter shapes.

4

4) Enable interfaces (optional, but recommended)

Enable interface responses. Add the book listing UI.

This example uses HandlebarsDisplay. That widget renders HTML from a template.

Training Agent UIs

5

5) Add the Handlebars template

Add a Handlebars template for the interface. The training template was generated by the 'Template Assistant' agent.

Agent Dynamic UI Template

Run it

After you Save, the agent appears in the agent list.

AI Agent Listing

Open the agent and try a prompt like:

• Find books by "Isaac Asimov".

• Show me the books with "The Hobbit" in title.

You should see tool-backed results. With interfaces enabled, you’ll get a rich HTML response.

AI Agent Interaction

Call it via API (optional)

Agents are also available as backend APIs. Use these endpoints from the runner channel that serves the agent:

• AI Agent APIs

Use the Call Path from the agent’s Definition tab. That value is the stable identifier for calling the agent remotely.

Troubleshooting

• Model can’t be called: the agent is not in Allowed For for your runner.

• LLM auth error: secret is missing or not resolved on the runner.

• Agent never uses tools: /train_rest is not listed in Tool Sagas.

• Tool call fails: verify /train_rest works directly first. Start from API Flow Examples.

• UI response is plain text: interfaces are disabled or no interface is assigned.

• Broken HTML: Handlebars template has invalid syntax. Simplify and retry.

Next steps

• Deep dive into model tabs and governance in GenAI Models.

• Extend the tool saga patterns in API Flow Examples.

• Build more admin screens like the training UI in UI Example.