Hands-on Lab571 words

Equip a travel-desk agent — decision exercise

Exercise: equip a travel-desk agent

Original fictional scenario. No cloud account, API calls or paid services are required. Difficulty: intermediate · Estimated duration: 15–20 minutes

Bluefin Travel, an invented corporate travel company, is designing an agent for its client desk. A contractor's draft proposes these tools:

NeedDraft proposal
1Answer questions from Bluefin's travel-policy pages and uploaded handbooks
2Book seats with an airline partner that publishes an OpenAPI schema
3Turn voicemails from travellers into text
4Translate itinerary PDFs while keeping their layout
5Extract totals and dates from scanned receipts
6Tag product videos with labels
7Connect the booking API through a new Vertex AI extension

The prototype was built by one developer in Google AI Studio on the free tier. Clients include a regulated bank.

Your decision

  1. Name the tool type or Google Cloud service for needs 1–6, and correct the draft where it is wrong.
  2. Explain what is wrong with need 7, and what to do instead.
  3. In function calling, who executes the booking tool, and why does that matter for permissions?
  4. Where should the production agent be built, and which two facts from Google's comparison decide it?

Rubric (10 house points)

  • 5 points: needs 1–5 correctly matched, with one reason each.
  • 1 point: need 6 flagged — Video Intelligence API shuts down in 2027; Gemini is recommended.
  • 2 points: need 7 replaced (extensions shut down after 26 November 2026) by an OpenAPI tool or function.
  • 2 points: production in Agent Studio on Agent Platform, citing data use and the regulated-customer guidance.

Reference solution

1 → a data store tool, which answers from website content and uploaded data. 2 → an OpenAPI tool built from the airline's schema. 3 → Speech-to-Text. 4 → Document Translation, which keeps the original formatting and layout. 5 → Document AI, which turns documents into structured fields.

Need 6: the Video Intelligence API can be used only until 14 September 2027; a new design should follow Google's recommendation to use Gemini models. Need 7: Vertex AI Extensions shuts down after 26 November 2026, so connect the booking API with the OpenAPI tool from need 2 or a function your application executes.

With function calling the model only outputs the call and its parameters; your application executes it, so permission checks belong in the application, not in the prompt.

Production belongs in Agent Studio on Gemini Enterprise Agent Platform: Google says regulated customers should use Agent Platform, where prompts, responses and data are never used to improve Google products — unlike the free tier the prototype ran on. Google AI Studio remains the right place for the fast prototype.

Rejected alternatives: plain text translation for need 4 (it loses layout); a data store tool for need 2 (it answers from content, it does not call a booking API); keeping the extension because it already works (it stops working after the shutdown date).

Sources and scope

All organisations, names, numbers, thresholds, the rubric and the conclusions above are house-authored. The product facts the reasoning depends on are grounded in:

Ready to study Generative AI Leader (GCP-GAIL)?

Practice tests, flashcards, and all study notes — free, no sign-up needed.

Start Studying — Free