AI Tech Group
TUI Concierge Lab · participant guide
0/0 steps
Google Cloud AI Tech Group Google Cloud × TUI

{{event}}

Build the TUI Holiday Concierge with agents

Build a live TUI travel concierge in Antigravity using nine ready-made skills. Gemini plans trips from our shared catalogue, Python checks every plan, and A2UI paints it onto a live canvas. Then extend it and show what you built.

50TUI destinations in the shared BigQuery graph
9tui-* agent skills with reference code
10,482place-to-place travel times in the graph
0mocked answers: every reply is live
1

Open Antigravity

Your Chromebook and ~/tui-build workspace are ready. Open Antigravity to get started.

Terminal
antigravity ~/tui-build
  1. Click Trust when asked. Antigravity picks up the nine skills from .agents/skills/.
  2. If it asks you to sign in: Business account › Continue with Google Cloud › pick your Gemini Enterprise seat (or Other › project {{project}}, global).
i

Use the Copy button on any box. Terminal commands go in the Chromebook Terminal; prompts go into the Antigravity chat. Lines starting with # aren't copied.

2

Build the concierge

Have Antigravity build it layer by layer, or run the finished reference app first.

~/tui-build already has the nine skills, the catalogue snapshot, and your .env. Paste this into Antigravity—it will build one layer at a time and pause after each smoke test until you reply go.

Starter promptAntigravity · ~/tui-build

Build the TUI Holiday Concierge

Build the TUI Holiday Concierge in this workspace using the tui-concierge-* skills: - TUI look and feel with a 3-column layout and a collapsible itinerary canvas on the right - Live chat that plans and edits trips using our shared BigQuery graph ({{project}}.{{dataset}}.{{graph}}, read-only) - Deterministic plan checks in Python and live A2UI canvas updates Follow tui-concierge-quickstart layer by layer, pause after each layer's smoke test so I can say "go", and start with layer 1.

Reply go after each layer. When it finishes, run make serve in ~/tui-build and open localhost:8080.

Follow-upif a layer fails

Fix a failing smoke test

That layer's smoke test failed. Compare it with the skill's reference files, fix what's off, and re-run the test.
i

Short on time? Build layers 1–5 yourself, then ask Antigravity to pull layers 6–7 from the reference. Layer 5 (planner + checker) is the heart of it.

{{repoDir}} has the complete working app—great for a quick preview or comparing against your build:

Terminal
cd {{repoDir}}
make test            # offline tests (~40 tests)
make serve           # http://localhost:8080

Open http://localhost:8080 and jump to Step 3. Press Ctrl+C to stop the server.

PromptAntigravity

How one chat turn works

Walk me through one traveller turn in this repo—from POST /api/chat/stream to the planner, plan checker, and A2UI response. Keep it short.
PromptAntigravity · optional

Quick overview (1 minute)

Skim tui-concierge-quickstart and give me a quick rundown of how the concierge works and the 7 build layers.
3

Try it like a traveller

With make serve running (your build or the reference), type these into the chat at localhost:8080.

TravellerConcierge chat

1 · Discover

Somewhere warm in early December for two, we love food and the sea.

Expect destination photo cards with temperature, flight time and "Plan this trip".

TravellerConcierge chat

2 · Plan

4 nights in Rome in late April for 2 adults from Gatwick. We love art and long lunches and hate crowds.

Watch "Behind the scenes" stream, the Colosseum draw itself into the header, and the plan appear with photos and a total.

TravellerConcierge chat

3 · Change it in your own words

Day 2 feels too touristy. Swap the busiest stop for something indoors and quieter, under £40 per person.

Expect a "Change applied" card and a yellow UPDATED stop. Only that part of the canvas is patched (A2UI updateDataModel).

TravellerConcierge chat

4 · Constraint that matters

My mother is joining us and can't walk more than 10 minutes at a time. Adjust the plan.

Then try the buttons

  1. Simulate event › Rain on day 2: the Trip Watcher changes only day 2, keeps the hotel and flights, and makes the 09:00–12:00 stops indoor.
  2. Undo restores the previous version. Map shows the stops on Google Maps. Governance shows the runtime, registry, data and A2UI wire.
  3. Switch the hotel in Your trip and watch the total change.
i

Curious what the canvas receives? Open Chrome DevTools › Network › stream and read the A2UI messages.

4

Extend it: pick a track

Pick one track for your team. Run make test after each change.

Track 1 · Data & graph

Indoor experiences near the hotel

A graph-powered tool the planner can call on rainy days.

Track 2 · Planner rules

A rule Gemini must respect

Add a Python check and watch the repair loop enforce it.

Track 3 · Trip Watcher

A new disruption

A local "museum closed" event, without touching BigQuery.

Track 4 · Canvas & design

Budget meter

A new A2UI widget in TUI's style, patched live.

Track 5 · Maps grounding

Walking routes on the map

Routes API v2 polylines between stops, key kept on the server.

Track 6 · Your idea

Pitch your own

Anything that makes the concierge better for TUI travellers.

Track 1tui-concierge-data-graph

Indoor experiences within walking distance

Using tui-concierge-data-graph, add a read-only helper find_indoor_near_hotel(destination_id, hotel_id, max_walk_mins=15) on {{graph}} (with an offline twin and test) and feed it to the planner when someone asks for rainy-day ideas.
Track 2tui-concierge-adk-replanner

At most two paid experiences per day

Using tui-concierge-adk-replanner, add a check in app/plan_checker.py so each day has at most two paid experiences and keeps at least 90 minutes free in the afternoon. Add a unit test and show it working live.
Track 3tui-concierge-adk-replanner

"Museum closed" disruption, local only

Using tui-concierge-adk-replanner, add a local "Museum closed on day 3" event (EVT_MUSEUM_CLOSED_DAY3, indoor replacement on day3_stop1) without writing to BigQuery, wire it into Simulate event, and add a test.
Track 4tui-concierge-design-system + a2ui-v1-canvas

Budget meter on the plan

Using tui-concierge-design-system and tui-concierge-a2ui-v1-canvas, add a budget meter under the plan header that updates live via A2UI when the traveller sets a budget (e.g. "our budget is £1,200") and turns amber above 90%.
Track 5tui-concierge-maps-grounding

Walking routes between stops

Using tui-concierge-maps-grounding, draw the walking or driving routes between each day's stops on the Map tab via a server-side /api/route/{day_number} endpoint (Routes API v2, key on server), with an offline test.
Track 6any skills

Shape your own idea

I want to add this to the concierge: <your idea>. Check tui-concierge-quickstart, tell me which files this touches, and suggest the smallest end-to-end change with a test before editing.
!

Keep the basics intact: Gemini only picks IDs, Python checks the plan, shared data stays read-only, and API keys stay on the server.

5

Show it

Five minutes per team: a live demo and a quick deck on how it works.

Prompttui-deck-gen

A TUI-branded readout deck

Using tui-deck-gen, build a 5-slide TUI-branded HTML deck about what our team built: title, traveller problem, 3-prompt live demo script, how it works under the hood, and what's next. Verify the layout and serve it on http://localhost:8000.
Terminal · preview the deck
# if the agent didn't start it for you
cd {{repoDir}}/.agents/skills/tui-deck-gen/examples && python3 -m http.server 8000

Demo checklist

  1. Start from a clean chat, plan a trip, then change it once.
  2. Show your new feature live and run make test.
  3. Call out which skills helped and one thing you had to steer.

Optional extras

Gemini Enterprise

Brainstorm test scenarios

Open Gemini Enterprise and ask: "Give me five tricky traveller prompts to stress-test a live holiday concierge (weather, mobility, budget, closures)."

Cloud Run

Deploy your own copy

Only if your facilitator enabled deploy rights. Use your own service name, never the shared tui-concierge-web:

Terminal · optional deploy
cd {{repoDir}}
SVC="tui-$(gcloud config get account | cut -d@ -f1)"
gcloud run deploy "$SVC" --source . --region europe-west1 --project {{project}} \
  --allow-unauthenticated --memory 2Gi --timeout 300 \
  --env-vars-file <(grep -E '^[A-Z_]+=' .env | sed -E 's/^([A-Z_]+)=(.*)$/\1: "\2"/')
★

Bonus: run an agent swarm

Got a few minutes left? Type /swarm in Antigravity to fan out parallel agents and QA your live app.

Make sure make serve is running on localhost:8080, then start a message with /swarm in Antigravity so multiple agents split the audit in parallel:

Bonus · /swarmAntigravity chat

Fact-check and click-test every Maps location

/swarm Audit our running concierge at http://localhost:8080 with parallel agents: - Fact-check trip details, prices, and timings against data/resolved_catalog.json - Click through every Google Maps location and place link in the itinerary and Map tab to verify every link loads and points to the right place - Summarize any broken links, mismatches, or console errors in a short table

Tip: you can aim /swarm at any big task—parallelising tests, reviewing UI states, or checking multiple destinations at once.

◆

The nine skills

Antigravity loads them from .agents/skills/. Each one owns a layer of the app.

SkillWhat it teaches the agent
tui-concierge-quickstartStart here: build order, setup, acceptance checklist, one-step rebuild (materialize.py)
tui-concierge-data-graphShared BigQuery catalogue + GQL on the property graph, offline snapshot
tui-concierge-maps-groundingPlaces API (New), Routes v2, server-side photo proxy, attribution rules
tui-concierge-a2ui-v1-canvasA2UI 1.0: createSurface once, then surgical updateDataModel patches
tui-concierge-adk-replannerLive planner: Gemini picks IDs, code checks the plan, repair loop, Trip Watcher rules, ADK agents
tui-concierge-registry-deployFastAPI server, Cloud Run, Agent Engine and Agent Registry governance
tui-concierge-brand-uiThe three-column front end: chat, streaming, photos, landmark banner, map
tui-concierge-design-systemTUI tokens, type, layout, components and reference screens (no Figma or Stitch needed)
tui-deck-genTUI-branded single-file HTML decks for your show-and-tell
PromptAntigravity

Query the live graph

Using tui-concierge-data-graph, run a read-only GQL query on {{project}}.{{dataset}}.{{graph}} for the 5 experiences closest to the Colosseum in Rome, with travel time and mode.
⌘

Command cheat sheet

Everything runs from {{repoDir}} in the Terminal.

CommandWhat it does
make doctorReadiness check (tools, sign-in, BigQuery, Gemini, Maps key)
make testOffline test suite (no BigQuery, stub model)
make test-liveLive BigQuery + Gemini end-to-end test
make serveRun the concierge on http://localhost:8080
make snapshotRefresh the offline copy of the shared catalogue (read-only)
open http://localhost:8080/static/specimen.htmlThe living design system (tokens, type, components)
Terminal · everyday loop
cd {{repoDir}}
make test && make serve
?

Troubleshooting

Most issues are one command away.

Antigravity keeps asking me to sign in, or says "keyring is locked"
Terminal
export $(dbus-launch)
antigravity {{repoDir}}
Antigravity says I have no licence or hit a quota

In the licence selector choose Other › project {{project}} › location global. If that fails too, tell your facilitator (they check your Gemini Enterprise seat and the agentspaceUser role).

"Not signed in", "Reauthentication is needed" or permission errors
Terminal
gcloud auth login --update-adc
gcloud config set project {{project}}
gcloud auth application-default set-quota-project {{project}}
The page at localhost:8080 doesn't load

Is make serve still running in a Terminal tab? Start it again. If Chrome still can't connect, try http://penguin.linux.test:8080.

No photos or no map (Maps key missing)

Your .env lacks the Maps key. Run this in the folder you serve from, then restart make serve:

Terminal
sed -i "s|^GOOGLE_MAPS_API_KEY=.*|GOOGLE_MAPS_API_KEY=$(gcloud secrets versions access latest --secret={{mapsSecret}} --project={{project}})|" .env
The agent wants to write to BigQuery or change the shared data

Say no. Other teams use the same tables; use local files (like data/local_events.json) or the offline snapshot with USE_LIVE_BQ=false.

Tests fail after a change
PromptAntigravity
make test failed. Run it, tell me what broke in a sentence or two, fix it cleanly, and re-run until green.
⚙

Facilitator setup

Done before the session, per Chromebook. Participants skip this.

  1. Sign in to ChromeOS as the participant (alpha / beta / gamma), then Settings › About ChromeOS › Developers › Linux development environment › Set up (20 GB or more).
  2. In the Terminal, run the installer. It installs Antigravity, the Google Cloud CLI and uv, selects project {{project}} (europe-west1), signs in once (covering both the CLI and application-default credentials), downloads the bundle and runs the readiness check.
Terminal · facilitator, once per Chromebook
curl -fsSL {{guideUrl}}/setup.sh | bash
cd {{repoDir}} && make doctor

Then open Antigravity once and sign in: Business account › Continue with Google Cloud › licence. Intel/AMD Chromebooks only: Antigravity crashes on ARM64.

§

Rules of the road

Short, and they matter.

  • Don't change the shared data. Your account could write to it, but all three teams use the same tables: put new data in local files or your own dataset. The catalogue is synthetic demo data built from real Google places; prices are indicative.
  • No secrets in prompts or code. The Maps key lives in .env and on the server only; never paste it into a chat or commit it.
  • Google Maps Platform terms: show the Google Maps attribution and photo author credits wherever a map or photo appears; use place data only inside this app; don't cache Places content beyond what the app already does.
  • Real answers only. No canned replies or seeded plans; the model chooses catalogue IDs and the plan checker has the final word.
  • Be kind to the shared project: one make serve at a time, stop long loops, and tell your facilitator if something burns quota.