エンドツーエンドでテストする

Run the quest classifier on three quests in the debugger and confirm it calls Monster Query only when a quest names a creature.

With the tool connected and the system prompt replaced, you can test the agent.

手順 5 - エンドツーエンドでテストする

1. Select Debug from the toolbar at the top of the agent builder.
2. Enter the following quest:

Slay the ancient red dragon terrorizing the countryside.

3. Select Save & Debug.

Watch the Execution Trail at the bottom of the agent builder. For this quest, expect:

1. The agent calls Monster Query with a search term for the ancient red dragon.
2. The tool returns one result: Ancient Red Dragon (srd-2024_ancient-red-dragon), challenge rating 24.
3. The agent returns a tier of Legendary, and its reasoning cites the challenge rating of 24.

Select the HTTP Request - Open5e Monster Query span to see the request the tool sent: the v2/creatures/ URL, the search term the agent chose in name__icontains, and the document__key and fields values you configured in Step 1.

Now run two more quests, letting each finish before you start the next:

Clear the rats out of the inn cellar. They are loud and they are eating all of the cheese.

Expected: one Monster Query call for rats, Swarm of Rats (challenge rating 0.25), tier Trivial.

Deliver a sealed letter to the magistrate in the next town.

Expected: no Monster Query call, because the quest names no creature, and tier Trivial.

The letter run matters as much as the dragon run. An agent that calls the tool on every quest wastes an API call and tokens each time, and an agent that never calls it is back to guessing. Step 7 measures both.

---

With the agent calling its tool only when it should, you can evaluate it in the next section.