75-Min GDE Workshop

Build a Location Intelligence
ADK Agent with MCP servers
for BigQuery and Google Maps

Hands-On Workshop • Ground a Gemini agent in enterprise data and the real world

Rabimba Karanjai • Google Developer Expert (AI/ML)

Module 0 • Welcome • 00:02 – 00:04 (2 min) Google Developer Expert

💳 Redeem Your Free Cloud Credits

g.dev/credits/latw-bigquery-100

Point your phone camera at the code → redeem → create or pick a project with billing enabled. BigQuery and Maps Grounding Lite are pay-as-you-go; today's lab fits inside these credits many times over.

Wi-Fi: LATW Dev Network · Password: LATWDevs2026!

Module 0 • Welcome • 00:04 – 00:05 (1 min) Google Developer Expert

This is what you'll have running in about 40 minutes

One Gemini agent • Two managed MCP servers • Zero hand-written API clients

Live demo: an ADK agent chains BigQuery SQL with Google Maps lookups (source: Google Cloud announcement, Dec 10 2025).

Module 0 • Welcome • 00:05 – 00:06 (1 min) Google Developer Expert

Workshop Hub — scan these now

Everything you need is one QR away; the codelab has all commands and prompts

</>

1. The Codelab (follow along)

codelabs.developers.google.com/adk-mcp-bigquery-maps

Full step-by-step lab: setup scripts, agent code, and every prompt we will run today.

</>

2. Free Credits (redeem first!)

g.dev/credits/latw-bigquery-100

Maps Grounding Lite and BigQuery need billing enabled. Redeem your event credit code, then create a project.

</>

3. Slides & Speaker Runbook

github.com/rabimba/GDE-ML-Artifacts

This deck, the verbatim facilitator script, and the copy-paste prompt sheet.

Google Developer Expert

Rabimba
Karanjai

Google Developer Expert — AI & ML

ML Scientist — PayPal

Agentic Systems Model Context Protocol Google Cloud Run

Module 0 • Welcome • 00:07 – 00:08 (1 min) Google Developer Expert

Start your engines — setup runs while I talk

1

Open Cloud Shell

console.cloud.google.com → Activate Cloud Shell (top bar)

2

Clone the demo repo

git clone https://github.com/google/mcp.git

3

Run setup_env.sh

cd mcp/examples/launchmybakery && ./setup/setup_env.sh

4

Run setup_bigquery.sh

./setup/setup_bigquery.sh

DO THIS NOW

Provision in the background

The two setup scripts take ~3-5 min. Kick them off now, then look up — I'll explain what everything is while your project provisions itself.

Cloud Shell

# Verify auth + project
gcloud auth list
export PROJECT_ID=$(gcloud config get project)

# Get the code
git clone https://github.com/google/mcp.git
cd mcp/examples/launchmybakery

# Kick off BOTH setup scripts
./setup/setup_env.sh
./setup/setup_bigquery.sh

Module 1 • Concepts • 00:08 – 00:10 (2 min) Google Developer Expert

Your friend is opening a sourdough
bakery in Los Angeles. Where?

🎤 Quick poll: how would you actually decide which LA zip code to open in?

A) Walk around and 'get a feel' for neighborhoods Click to Vote
B) Hire a market-research firm ($$$, weeks) Click to Vote
C) Ask an LLM chatbot for the 'best' zip code Click to Vote
D) Combine real foot-traffic data with live Maps info Click to Vote

💡 C alone hallucinates — a model can't know current foot traffic or which bakeries exist. The winning move is D: data + the real world, joined by an agent. Today we build exactly that.

Module 1 • Concepts • 00:10 – 00:11 (1 min) Google Developer Expert

Two data worlds, one question

Enterprise Data — BigQuery

Demographics & foot-traffic index by zip
Competitor product pricing
Weekly sales history per store
Your private, governed numbers

An agent must reason across BOTH to answer 'where should I open?'

The Real World — Google Maps

Which bakeries actually exist nearby?
Drive time to the supplier
Weather for opening weekend
Fresh, public, geospatial truth

Module 1 • Concepts • 00:11 – 00:13 (2 min) Google Developer Expert

What makes an LLM an Agent?

Build the definition up piece by piece — press → to reveal each part

1. MODEL

Gemini 3.5 Flash — the reasoning brain. Alone, it only predicts text.

2. + INSTRUCTIONS

A system prompt that defines the goal, persona, and rules of the game.

3. + TOOLS

Functions the model may call to act on the world — query BigQuery, search Maps.

4. + LOOP

Reason → pick a tool → act → observe result → reason again, until the goal is met.

Agent = Model + Instructions + Tools, running in a Reason-Act-Observe loop

Module 1 • Concepts • 00:13 – 00:14 (1 min) Google Developer Expert

The three pillars we'll use today

Gemini 3.5 Flash
(on Vertex AI)

The reasoning model. Our .env sets GOOGLE_GENAI_USE_VERTEXAI=1, so ADK calls it through your project's Vertex AI endpoint.

ADK
Agent Development Kit

Google's open-source framework for building agents: LlmAgent, tool wiring, sessions, and a built-in web UI for debugging.

MCP
Model Context Protocol

The open standard (started by Anthropic) that lets any agent discover and call tools on any server — one protocol instead of N custom integrations.

Module 1 • Concepts • 00:14 – 00:15 (1 min) Google Developer Expert

An ADK agent in ~15 lines

adk_agent/mcp_bakery_app/agent.py — the entire agent definition

python

root_agent = LlmAgent(
    model='gemini-3.5-flash',
    name='root_agent',
    instruction=f"""
      Help the user answer questions by strategically
      combining insights from two sources:
      1. **BigQuery toolset:** Access demographic (inc.
         foot traffic index), product pricing, and
         historical sales data in the mcp_bakery
         dataset. Do not use any other dataset.
         Run all query jobs from project id: {PROJECT_ID}.
      2. **Maps Toolset:** Use this for real-world
         location analysis, finding competition/places
         and calculating necessary travel routes.
         Include a hyperlink to an interactive map in
         your response where appropriate.
    """,
    tools=[maps_toolset, bigquery_toolset]
)

The three control surfaces:

model — Which Gemini does the thinking
instruction — The agent's job description — notice it scopes the data AND names the project
tools — Two MCP toolsets = every tool from both servers

Module 1 • Concepts • 00:15 – 00:17 (2 min) Google Developer Expert

MCP: from N×M integrations to N+M

The 'USB-C for AI' — Anthropic's open standard, adopted across the industry

Without MCP: every agent hand-codes every API

ADK Agent

Gemini CLI

Any MCP client

✕ ✕ ✕
✕ ✕ ✕

BigQuery

Google Maps

Your API?

With MCP: each side implements the protocol once

ADK Agent

Gemini CLI

Any MCP client

MCP

BigQuery

Google Maps

Your API?

Module 1 • Concepts • 00:17 – 00:18 (1 min) Google Developer Expert

MCP anatomy: Host, Client, Server

And the transport that carries it — today that's Streamable HTTP

HOST

ADK's web UI — the app the user chats with.

CLIENT

An MCP client inside the host keeps a dedicated connection per server.

SERVER

BigQuery & Maps servers expose a catalog of named, typed tools.

TRANSPORT

Stdio for local servers; Streamable HTTP for remote, Google-hosted ones — what we use today.

Our agent never runs BigQuery or Maps code — it calls remote tools over HTTP

Module 1 • Concepts • 00:18 – 00:20 (2 min) Google Developer Expert

Google's managed, remote MCP servers

Announced Dec 10, 2025 — Google's API infrastructure now speaks MCP

No local servers to install, patch, or secure. You point your client at a globally-consistent endpoint.

BigQuery: https://bigquery.googleapis.com/mcp
Google Maps (Grounding Lite): https://mapstools.googleapis.com/mcp
Also live: GCE & GKE today — rolling out next: Spanner, AlloyDB, Cloud Run, Developer Knowledge...
Enable per project: gcloud beta services mcp enable <service> (setup_env.sh did this)
Governance: IAM (roles/mcp.toolUser), audit logging, Model Armor for prompt-injection defense

🔍 Click diagram to enlarge full-screen

Module 1 • Concepts • 00:20 – 00:21 (1 min) Google Developer Expert

BigQuery MCP server — the 6 tools

The agent discovers these itself; column descriptions become its documentation

list_dataset_idsWhat datasets exist in this project?

list_table_idsWhat tables are in mcp_bakery?

get_dataset_infoMetadata about a dataset

get_table_infoSchema + column descriptions for a table

execute_sqlRun any SQL — including AI.FORECAST

execute_sql_readonlySame, but mutations blocked by design

The agent reads get_table_info output to write its own SQL. Rich column descriptions in setup_bigquery.sh = a better agent. Your schema docs ARE the prompt.

Module 1 • Concepts • 00:21 – 00:22 (1 min) Google Developer Expert

Maps Grounding Lite — the 5 tools

Trusted geospatial grounding with required Google Maps attribution

search_placesFind places + AI summaries + Maps links

compute_routesDistance & duration between two points

lookup_weatherCurrent conditions & forecasts

resolve_namesFree-text address/place -> Place ID

resolve_maps_urlsGoogle Maps URL -> Place ID

Grounding Lite returns distance & duration — NOT turn-by-turn directions. Every grounded answer must show its Google Maps source link (the agent is instructed to include it).

Module 1 • Concepts • 00:22 – 00:23 (1 min) Google Developer Expert

Two servers, two auth patterns — tools.py

Same MCPToolset, different credentials

Both toolsets use StreamableHTTPConnectionParams; only the header differs.

Maps: Simple API key in the X-Goog-Api-Key header (restricted to mapstools.googleapis.com).
BigQuery: OAuth 2.0 bearer token from your Application Default Credentials + x-goog-user-project for billing.
⚠️ Token lifetime: ADC tokens expire ~60 min and ADK does not auto-refresh — re-run gcloud auth application-default login if a long session dies.

adk_agent/mcp_bakery_app/tools.py (excerpt)

# --- Maps: API key header ---
MCPToolset(
  connection_params=StreamableHTTPConnectionParams(
    url="https://mapstools.googleapis.com/mcp",
    headers={"X-Goog-Api-Key": maps_api_key},
    timeout=30.0, sse_read_timeout=300.0))

# --- BigQuery: OAuth bearer from ADC ---
credentials, project_id = google.auth.default(
    scopes=["https://www.googleapis.com/auth/bigquery"])
credentials.refresh(
    google.auth.transport.requests.Request())
HEADERS_WITH_OAUTH = {
    "Authorization": f"Bearer {credentials.token}",
    "x-goog-user-project": project_id}
MCPToolset(
  connection_params=StreamableHTTPConnectionParams(
    url="https://bigquery.googleapis.com/mcp",
    headers=HEADERS_WITH_OAUTH,
    timeout=30.0, sse_read_timeout=300.0))

Module 1 • Concepts • 00:23 – 00:25 (2 min) Google Developer Expert

The reasoning journey: macro → micro → money

Watch how one conversation walks down the decision funnel, switching tools

Q1 Which zip has the highest morning foot traffic? BigQuery • execute_sql → 90403 (Santa Monica)

Q2 Is 90403 already saturated with bakeries? Maps • search_places → competitor scan

Q3 Max price of a Sourdough Loaf in LA Metro? BigQuery • execute_sql → Erewhon ≈ $18.50

Q4 Forecast Dec sales from my best store @ $18 BigQuery • AI.FORECAST → revenue projection

Q5 Is Restaurant Depot < 30 min away? Maps • compute_routes → logistics check

Module 1 • Concepts • 00:25 – 00:26 (1 min) Google Developer Expert

The full architecture

Everything on one slide — this is your 'you are here' map for the lab

Trace one request end to end:

1. User: asks in the ADK web UI (adk web)
2. Agent (Gemini 3.5 Flash): reasons on Vertex AI, picks a tool
3. MCP toolset: sends a signed Streamable HTTP call
4a. BigQuery MCP: interprets schema & runs SQL in place — data never leaves BigQuery
4b. Maps MCP: grounds in live places, routes, weather
5. Agent: combines both into one grounded business answer

🔍 Click diagram to enlarge full-screen

Module 1 • Concepts • 00:26 – 00:27 (1 min) Google Developer Expert

Concept Check

Click to reveal each answer

Q1: Agent needs today's weather at the site?
Q2: Agent must forecast December revenue?
Q3: Agent must list tables before writing SQL?

Why not just…?

Common objections

Paste CSVs into the prompt?: 40k+ rows won't fit, goes stale, and leaves governance.
Call Maps with curl?: You'd hand-write auth, parsing & schemas — MCP already did it.
Fine-tune on my data?: Teaches style, not fresh facts. Tools > training for live data.

Module 1 • Concepts • 00:27 – 00:28 (1 min) Google Developer Expert

Taking this to production — the guardrails that matter

The demo uses your personal credentials for speed. Before this pattern ships, layer on:

• Least privilege: grant only roles/mcp.toolUser + bigquery.jobUser + bigquery.dataViewer; nothing broader
• Separate agent identity: run the agent as its own service account so tool calls are attributable & revocable
• Prefer read-only SQL: use execute_sql_readonly, or an IAM deny policy on execute_sql, so the agent can't mutate
• Restrict the Maps key: scope it to mapstools.googleapis.com only (setup_env.sh already did)
• Defend the prompt: Model Armor screens tool output for indirect prompt injection
• Watch the logs: every MCP tool call lands in Cloud Audit Logs — alert on anomalies

Module 2 • Lab A: Setup • 00:28 – 00:29 (1 min) Google Developer Expert

Let's build this!

Part A: finish provisioning • Part B: run the agent • Part C: the full journey

Module 2 • Lab A: Setup • 00:29 – 00:31 (2 min) Google Developer Expert

Step 1: Project, auth & the code

If you did the 'engines' step, just verify — else run it now

Confirm you're authenticated and the repo is cloned.

gcloud auth list: must show your account as ACTIVE
export PROJECT_ID=$(gcloud config get project): every later script reads this
gcloud auth application-default login: creates the ADC the BigQuery MCP toolset needs
⚠️ Billing: Maps Grounding Lite & BigQuery both require billing — redeem your credit first

Cloud Shell

gcloud auth list
gcloud config get project
export PROJECT_ID=$(gcloud config get project)

git clone https://github.com/google/mcp.git
cd mcp/examples/launchmybakery

# Required so ADK can call BigQuery via MCP
gcloud auth application-default login

🔍 Click screenshot to zoom

Module 2 • Lab A: Setup • 00:31 – 00:33 (2 min) Google Developer Expert

What those two setup scripts just did

Press → to walk through the provisioning you kicked off at minute 5

setup_env.sh • 1

Enables APIs: aiplatform, apikeys, mapstools, bigquery — plus the MCP endpoints (gcloud beta services mcp enable).

setup_env.sh • 2

Creates a Maps API key restricted to mapstools.googleapis.com and writes .env (VERTEXAI flag, project, key).

setup_bigquery.sh • 1

Creates a Cloud Storage bucket and uploads the 4 CSVs (demographics, prices, sales, foot traffic).

setup_bigquery.sh • 2

Creates dataset mcp_bakery and 4 tables — each column carrying a description the agent will read.

Result: a governed dataset + a restricted key + MCP endpoints switched on

Module 2 • Lab A: Setup • 00:33 – 00:35 (2 min) Google Developer Expert

✋ Checkpoint 1: Your data is live

Open: the BigQuery console in your project
Confirm: dataset mcp_bakery exists with 4 tables
Peek: at foot_traffic → rows where time_of_day='morning' carry a foot_traffic_score
Stuck?: re-run ./setup/setup_bigquery.sh; check PROJECT_ID is set

BigQuery console: mcp_bakery dataset with its 4 tables (Click to zoom)

Module 2 • Lab A: Setup • 00:35 – 00:38 (3 min) Google Developer Expert

Step 2: Install ADK in a virtualenv

google-adk[mcp] — the [mcp] extra brings the MCP client

Isolate the install, then install ADK with MCP support.

python3 -m venv .venv: keep Cloud Shell's system Python clean
pip install "google-adk[mcp]": the extra pulls in the MCP client libraries
Trouble?: the repo pins a known-good: pip install google-adk==1.28.0
Then: cd adk_agent/ — adk web discovers apps under the CWD

Cloud Shell

python3 -m venv .venv
source .venv/bin/activate

# The [mcp] extra = MCP client support
pip install "google-adk[mcp]"

# adk web scans the current dir for agents
cd adk_agent/

Module 3 • Lab B: Run It • 00:38 – 00:40 (2 min) Google Developer Expert

Step 3: Read the agent you'll run

Open Editor: in Cloud Shell, browse mcp/examples/launchmybakery
agent.py: find the LlmAgent — model, instruction, tools
tools.py: find the two MCPToolset calls & their headers
.env: confirm GOOGLE_GENAI_USE_VERTEXAI=1 and MAPS_API_KEY are set

Cloud Shell Editor showing the launchmybakery layout (Click to zoom)

Module 3 • Lab B: Run It • 00:40 – 00:43 (3 min) Google Developer Expert

Step 4: Launch the agent UI

adk web + Cloud Shell Web Preview on port 8000

Start the dev server, then open it through Web Preview.

--allow_origins: required in Cloud Shell so the IDE proxy origin is accepted
Web Preview: top-right of Cloud Shell → Change port → 8000 → Change and Preview
Pick the app: select mcp_bakery_app in the UI's dropdown
Empty dropdown?: you ran adk web from the wrong directory — cd adk_agent/ first

Cloud Shell (inside adk_agent/)

adk web --allow_origins \
  'regex:https://.*\.cloudshell\.dev'

# Then in Cloud Shell:
#  Web Preview (top right)
#    → Change port → 8000
#    → Change and Preview

🔍 Click screenshot to zoom

Module 3 • Lab B: Run It • 00:43 – 00:46 (3 min) Google Developer Expert

✋ Checkpoint 2: First tool call (macro)

Ask: "I want to open a bakery in Los Angeles. Find the zip code with the highest morning foot traffic score."
Watch the trace: agent calls get_table_info, then execute_sql on foot_traffic
Expect: zip 90403 (Santa Monica) surfaces to the top
Auth error?: re-run gcloud auth application-default login (60-min token)

ADK web: execute_sql finds the top morning-traffic zip (Click to zoom)

Module 3 • Lab B: Run It • 00:46 – 00:49 (3 min) Google Developer Expert

✋ Checkpoint 3: Validate on the ground (micro)

Follow up: "Can you search for 'Bakeries' in that zip code to see if it's saturated?"
Notice: no zip in your prompt — the agent carried 90403 from memory
Watch: the trace switch servers: Maps search_places
Click: the Google Maps link in the answer — that's the required attribution

ADK web: Maps search_places scans 90403 for bakeries (Click to zoom)

Module 3 • Lab B: Run It • 00:49 – 00:52 (3 min) Google Developer Expert

Your turn: run the rest of the journey

Three more prompts — predict the tool before you run each (answers in the codelab)

Q3 Max price of a Sourdough Loaf in LA Metro? BigQuery Erewhon ≈ $18.50

Q4 Forecast Dec '25 from best store, price $18 BigQuery + AI.FORECAST qty → ×$18 revenue

Q5 Closest Restaurant Depot under 30 min away? Maps compute_routes check

★ Chain them yourself — one thread, all five both servers full business case

Module 3 • Lab B: Run It • 00:52 – 00:54 (2 min) Google Developer Expert

⚡ Finished early? Break it (productively)

Tweak the instruction

The cheapest way to change behavior

Prompt / Snippet

# agent.py — edit the instruction, e.g.:
"Position the brand as budget-friendly
 and always cite the source table."
# save → adk web reloads → re-ask Q3

Watch: the price recommendation shift with one sentence
Try: a persona: 'You are a skeptical CFO'

Stretch the tools

Push past the script

Prompt / Snippet

# Ask about the weather on opening day
"Weather in 90403 this Saturday?"
   → Maps lookup_weather

# Safety: force read-only SQL
# MCPToolset(tool_filter=['execute_sql_readonly'])
# or an IAM deny on execute_sql

Weather: a 3rd tool you never configured — it was just there
Read-only: one change removes the agent's ability to mutate data

Module 3 • Lab B: Run It • 00:54 – 00:57 (3 min) Google Developer Expert

Quick fixes — the 6 ways today breaks

Hit a wall? It's almost certainly one of these:

• Auth errors mid-lab: ADC token expired (~60 min) → gcloud auth application-default login
• 'API not enabled' / 403: MCP endpoint off → gcloud beta services mcp enable bigquery.googleapis.com && gcloud beta services mcp enable mapstools.googleapis.com
• Empty app dropdown: adk web run from wrong dir → cd mcp/examples/launchmybakery/adk_agent
• Web Preview refuses: missing --allow_origins 'regex:https://.*\.cloudshell\.dev'
• Maps calls fail: key missing/restricted wrong → check .env MAPS_API_KEY, target service mapstools.googleapis.com
• Billing errors: project has no billing → redeem the event credit and re-attach

Module 4 • Wrap-Up • 00:57 – 00:59 (2 min) Google Developer Expert

Clean up — don't donate your credits back

One script deletes the dataset, bucket, and API key

Stop the server (Ctrl+C), then run cleanup from the repo root.

Deletes: BigQuery dataset mcp_bakery + the GCS bucket
Revokes: the restricted Maps API key created at setup
Keeps: your project — so you can re-run ./setup scripts any time
Why it matters: BigQuery storage & Maps calls are pay-as-you-go

Cloud Shell (repo root)

# Ctrl+C to stop adk web, then:
cd ~/mcp/examples/launchmybakery

chmod +x cleanup/cleanup_env.sh
./cleanup/cleanup_env.sh

# Deletes: mcp_bakery dataset,
#  the staging bucket, and the API key.

Module 4 • Wrap-Up • 00:59 – 01:02 (3 min) Google Developer Expert

Your Location-Intelligence Agent playbook

The four ideas to take with you

Everything today reduces to four moves you can reuse for any agent + any data.

1

Ground,
don't guess

An LLM alone hallucinates business facts. Ground it in your warehouse and live Maps data.

2

One agent,
many tools

A single ADK LlmAgent + MCP toolsets reasons across BigQuery AND Maps — no glue code.

3

Schema is
the prompt

get_table_info feeds column descriptions to the model. Documented data = smarter agent.

4

Govern by
default

Managed MCP servers bring IAM, audit logs, read-only SQL, and Model Armor. Ship with guardrails.

Module 4 • Wrap-Up • 01:02 – 01:04 (2 min) Google Developer Expert

Where to take this next

Deploy the agent

Move off adk web: containerize and ship to Cloud Run or Agent Engine, with a service-account identity.

More MCP servers

Live now: GCE & GKE. Rolling out: Spanner, AlloyDB, Cloud Run, Developer Knowledge — full list: google/mcp on GitHub.

Your own data

Point the agent at YOUR dataset: write rich column descriptions, scope the instruction, start with execute_sql_readonly.

Mission Complete 🎉

Build a Location Intelligence
ADK Agent with MCP servers
for BigQuery and Google Maps

You built it: one grounded agent over your data and the real world

Rabimba Karanjai • Google Developer Expert (AI/ML)

Google Developer Expert

Thank you!

Questions & live troubleshooting.
The codelab stays free — take it, fork it, teach it.

Codelab adk-mcp-bigquery-maps

Connect on LinkedIn in/rabimba-karanjai

Slides & Runbook GDE-ML-Artifacts

Appendix • Reference • Reference Slide Google Developer Expert

Cheat sheet — every command from today

Setup

Run once per project

Prompt / Snippet

export PROJECT_ID=$(gcloud config get project)
git clone https://github.com/google/mcp.git
cd mcp/examples/launchmybakery
gcloud auth application-default login
./setup/setup_env.sh
./setup/setup_bigquery.sh

MCP endpoints: gcloud beta services mcp enable <svc>
Billing: required for BigQuery + Maps

Run & clean up

Each session

Prompt / Snippet

python3 -m venv .venv && source .venv/bin/activate
pip install "google-adk[mcp]"
cd adk_agent/
adk web --allow_origins \
  'regex:https://.*\.cloudshell\.dev'

# later, from the repo root:
cd ~/mcp/examples/launchmybakery
chmod +x cleanup/cleanup_env.sh
./cleanup/cleanup_env.sh

Token: re-login after ~60 min
Read-only: prefer execute_sql_readonly in prod

Build a Location IntelligenceADK Agent with MCP serversfor BigQuery and Google Maps

💳 Redeem Your Free Cloud Credits

This is what you'll have running in about 40 minutes

Workshop Hub — scan these now

1. The Codelab (follow along)

2. Free Credits (redeem first!)

3. Slides & Speaker Runbook

RabimbaKaranjai

Start your engines — setup runs while I talk

Open Cloud Shell

Clone the demo repo

Run setup_env.sh

Run setup_bigquery.sh

Provision in the background

Your friend is opening a sourdoughbakery in Los Angeles. Where?

Two data worlds, one question

Enterprise Data — BigQuery

The Real World — Google Maps

What makes an LLM an Agent?

1. MODEL

2. + INSTRUCTIONS

3. + TOOLS

4. + LOOP

The three pillars we'll use today

Gemini 3.5 Flash(on Vertex AI)

ADKAgent Development Kit

MCPModel Context Protocol

An ADK agent in ~15 lines

adk_agent/mcp_bakery_app/agent.py — the entire agent definition

MCP: from N×M integrations to N+M

MCP anatomy: Host, Client, Server

HOST

CLIENT

SERVER

TRANSPORT

Google's managed, remote MCP servers

Announced Dec 10, 2025 — Google's API infrastructure now speaks MCP

BigQuery MCP server — the 6 tools

Maps Grounding Lite — the 5 tools

Two servers, two auth patterns — tools.py

Same MCPToolset, different credentials

The reasoning journey: macro → micro → money

The full architecture

Everything on one slide — this is your 'you are here' map for the lab

Concept Check

Click to reveal each answer

Why not just…?

Common objections

Taking this to production — the guardrails that matter

Let's build this!

Step 1: Project, auth & the code

If you did the 'engines' step, just verify — else run it now

What those two setup scripts just did

setup_env.sh • 1

setup_env.sh • 2

setup_bigquery.sh • 1

setup_bigquery.sh • 2

✋ Checkpoint 1: Your data is live

Step 2: Install ADK in a virtualenv

google-adk[mcp] — the [mcp] extra brings the MCP client

Step 3: Read the agent you'll run

Step 4: Launch the agent UI

adk web + Cloud Shell Web Preview on port 8000

✋ Checkpoint 2: First tool call (macro)

✋ Checkpoint 3: Validate on the ground (micro)

Your turn: run the rest of the journey

⚡ Finished early? Break it (productively)

Tweak the instruction

The cheapest way to change behavior

Stretch the tools

Push past the script

Quick fixes — the 6 ways today breaks

Clean up — don't donate your credits back

One script deletes the dataset, bucket, and API key

Your Location-Intelligence Agent playbook

The four ideas to take with you

Ground,don't guess

One agent,many tools

Schema isthe prompt

Govern bydefault

Build a Location Intelligence
ADK Agent with MCP servers
for BigQuery and Google Maps

Rabimba
Karanjai

Your friend is opening a sourdough
bakery in Los Angeles. Where?

Gemini 3.5 Flash
(on Vertex AI)

ADK
Agent Development Kit

MCP
Model Context Protocol

Ground,
don't guess

One agent,
many tools

Schema is
the prompt

Govern by
default

Build a Location Intelligence
ADK Agent with MCP servers
for BigQuery and Google Maps