75-Min Interactive Workshop

Command and control:
Orchestrate app development
with Gemini and MCP

Hands-On Workshop • Build & Deploy Full-Stack Agentic Apps with Zero Manual Code

Rabimba Karanjai • Google Developer Expert (AI/ML)

Module 0 • Welcome & Onboarding • 00:02 – 00:04 (2 min) Proprietary + Confidential

💳 Redeem Your Free Cloud Credits

g.dev/credits/latw-mcp-010

Point your phone camera at the code → redeem → create or pick a project with billing enabled. We deploy to Cloud Run at the end of today's workshop, and billing must be on for that to work.

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

Module 0 • Welcome & Onboarding • 00:04 – 00:05 (1 min) Proprietary + Confidential

Workshop Hub & Essential Links

Scan or open the Codelab now so you can follow along step-by-step

</>

1. Workshop Codelab (Start Here)

codelabs.developers.google.com/gemini-mcp-agy

Full step-by-step instructions, prompts, and Cloud Shell setup for today's lab.

</>

2. Free Credits (redeem first!)

g.dev/credits/latw-mcp-010

Today's lab deploys to Cloud Run — redeem your event credit, then create a project with billing on.

</>

3. Google Antigravity Docs

antigravity.google/docs

Official documentation for AGY CLI, IDE, MCP servers, Skills, Plugins, and Hooks.

</>

4. Slides, Runbook & GDE Artifacts

github.com/rabimba/GDE-ML-Artifacts

Today's slides, copy-paste prompt sheet, and Antigravity SDK reference code.

Proprietary + Confidential

Rabimba
Karanjai

Google Developer Expert — AI & ML

ML Scientist — PayPal

Agentic Systems Model Context Protocol Google Cloud Run

Module 0 • Welcome & Onboarding • 00:07 – 00:10 (3 min) Proprietary + Confidential

Today's 75-Minute Mission

How we'll go from an empty Cloud Shell to a deployed production agent with zero manual coding

We will build the "Cat Delivery System" through 4 progressive hands-on checkpoints. Every module pairs a core agentic engineering pattern with immediate hands-on execution.

1

Hello AGY CLI
(00:20 – 00:30)

Launch & authenticate Antigravity CLI inside Cloud Shell Editor. Generate, run, and inspect our first Node.js cat trivia script.

2

Developer Knowledge MCP
(00:30 – 00:45)

Eliminate documentation cutoffs! Connect Google's official MCP server and generate kitten art with Nano Banana 2.

3

Skills & /plan Mode
(00:45 – 01:02)

Install the Antigravity SDK Skill & Chrome Web Plugin. Use /plan and /artifact to architect the Cat Delivery System.

4

Cloud Run & Guardrails
(01:02 – 01:10)

Switch to /fast mode to containerize and deploy to Cloud Run, then try to break our 'Strictly No Dogs (Capybaras OK)' guardrail!

Module 1 • Concepts & Architecture • 00:10 – 00:12 (2 min) Proprietary + Confidential

From Code Autocomplete
to Command & Control

🎤 Quick Show of Hands: How many of you have asked an AI coding assistant to use a brand-new SDK or model, and watched it confidently invent a method that doesn't exist?

🙋 Every single week! (The 'Hallucinated SDK' trap) Click to Vote
🙋 It used an outdated v0.1 API from 2023! Click to Vote
🙋 I ended up copy-pasting 10 pages of docs into the chat prompt! Click to Vote

💡 Why does this happen? Because even frontier LLMs have static training cutoffs and no native awareness of your team's architecture—unless we give them authoritative tools (MCP) and structured workflows (Skills).

Module 1 • Concepts & Architecture • 00:12 – 00:14 (2 min) Proprietary + Confidential

An AI coding agent is a system that uses a model to reason, an agent harness to plan & execute, and MCP & skills to access authoritative knowledge and tools.

MCP & Skills
(Context & Tools)

⟷

Antigravity
(Agent Harness)

⟷

Gemini Models
(Reasoning & Media)

Module 1 • Concepts & Architecture • 00:14 – 00:15 (1 min) Proprietary + Confidential

Objectives & Workshop Outcomes

By the end of this hands-on workshop, you will orchestrate Antigravity to build and deploy the "Cat Delivery System" while mastering four modern agentic development skills:

• Dual-Mode Orchestration (/fast vs /plan): Learn when to use /fast mode for immediate scripts and deployments versus /plan mode with interactive /artifact reviews for full-stack architectures.
• Authoritative Grounding via Developer Knowledge MCP: Connect Antigravity CLI directly to Google's official documentation server (developerknowledge.googleapis.com/mcp) so it writes up-to-date Vertex AI code on the first try.
• Progressive Disclosure with Agent Skills & Plugins: Install specialized skills (google-antigravity-sdk) and plugins (GoogleChrome/modern-web-guidance) that load into context only when needed.
• End-to-End Cloud Run Deployment & Auditing: Package a Lit + Vite frontend and FastAPI + Antigravity SDK backend into a single container, deploy to Cloud Run, and verify acceptance criteria automatically.

Module 1 • Concepts & Architecture • 00:15 – 00:17 (2 min) Proprietary + Confidential

Core Architecture

Flow of the "Cat Delivery System" — An Agent Building an Agent

Today's workshop features a two-tier agentic architecture: we use Antigravity CLI (Tier 1) to build and deploy a Cloud Run application powered by the Antigravity SDK (Tier 2).

Orchestration Layer (AGY CLI): Runs in Cloud Shell Editor, combining Developer Knowledge MCP + installed Skills to plan, write, and deploy the app.
Web Frontend (Lit + Vite): Clean, light-mode reactive chat interface bundled and served statically by our Python backend.
Agent Runtime (FastAPI + AGY SDK): Uses Gemini 3.5 Flash as the root agent with strict persona instructions: ABSOLUTELY NO DOGS (Capybaras allowed!).
Image Generation (Nano Banana 2): Invokes Gemini 3.1 Flash Image on Vertex AI to generate anime, 3D, or photorealistic cats on demand.

🔍 Click diagram to enlarge full-screen

Module 1 • Concepts & Architecture • 00:17 – 00:18 (1 min) Proprietary + Confidential

Key Technologies & Concepts We'll Use!

Google Antigravity
(CLI & Python SDK)

Google's AI-first agentic development platform. We'll use the terminal CLI (`agy`) to orchestrate code, and the Python SDK (`google-antigravity`) as the runtime harness for our app.

Developer Knowledge
MCP & Agent Skills

Open standards for extending agents. MCP connects `agy` to live Google documentation APIs, while Agent Skills (`SKILL.md`) provide progressive, on-demand procedural expertise.

Gemini 3.5 Flash, Nano
Banana 2 & Cloud Run

Gemini 3.5 Flash powers fast agent reasoning, Gemini 3.1 Flash Image (Nano Banana 2) generates images, and Cloud Run hosts our container serverlessly.

Module 1 • Concepts & Architecture • 00:18 – 00:20 (2 min) Proprietary + Confidential

Model Context Protocol (MCP)

Live External Tools & Authoritative Data

What It Is: An open protocol connecting AI agents to external APIs, databases, and documentation servers via standardized tool calls.
Google Developer Knowledge MCP: Hosted at developerknowledge.googleapis.com/mcp. Lets coding agents search & fetch live docs for Cloud, Vertex AI, Firebase, Android, and Maps.
Best For: Dynamic lookups, querying live systems, and ensuring SDK/API usage is 100% current.

Agent Skills & Plugins

Progressive Disclosure of Workflows (agentskills.io)

What They Are: Lightweight folders containing a SKILL.md instruction file (plus optional scripts/templates) that teach an agent how to perform specialized tasks.
Progressive Disclosure: At startup, ONLY the skill's name & short description are in the prompt. The full SKILL.md is loaded ONLY when the agent activates it—saving massive tokens!
Plugins: Bundle Skills + MCP servers + Rules + Hooks into a single installable package.

Module 2 • Lab Step 1: Hello Antigravity CLI • 00:20 – 00:21 (1 min) Proprietary + Confidential

Let's build this!

Lab Step 1 — Setting Up Cloud Shell Editor & Authenticating Antigravity CLI (agy)

Module 2 • Lab Step 1: Hello Antigravity CLI • 00:21 – 00:24 (3 min) Proprietary + Confidential

Step 1: Open Cloud Shell Editor Terminal

Critical Workaround for Clean TUI Rendering

While Cloud Shell comes with Antigravity pre-installed, the default Cloud Shell bottom pane doesn't render advanced TUI menus properly. Always use the terminal inside Cloud Shell Editor!

1. Activate Cloud Shell: Click the 'Activate Cloud Shell' icon (>_) in the top-right toolbar of the Google Cloud Console.
2. Click 'Open Editor': Click the pencil/editor icon on the Cloud Shell toolbar (and maximize the window) to launch the Cloud Shell Editor.
3. Open Integrated Terminal: Use menu Terminal > New Terminal or press Ctrl + ` (Control + backtick) to open the bash prompt inside the editor.

terminal — Cloud Shell Editor (Ctrl + `)

# Create a working directory for our warm-up exercise
mkdir -p ~/hello-agy && cd ~/hello-agy

# Launch Antigravity CLI
agy

🔍 Click screenshot to zoom

Module 2 • Lab Step 1: Hello Antigravity CLI • 00:24 – 00:27 (3 min) Proprietary + Confidential

Step 2: Authenticate Antigravity CLI

01

Select Auth Mode

Choose 'Use a Google Cloud project'

02

Complete OAuth Flow

Open link & paste the auth code

03

Set Project & Location

Enter your PROJECT_ID & choose 'global'

AUTHENTICATION CHECKLIST

Your Own Account & 'global' Location

Sign in with the same Google account that redeemed today's credit code. Already signed into a work org account? Open the OAuth link in an incognito tab instead. When prompted for location, select 'global'.

agy — First-Time Authentication Flow

? How would you like to authenticate?
> 1. Use a Google Cloud project

Paste authorization code: [PASTE_CODE_HERE]
Enter Google Cloud Project ID: [YOUR_PROJECT_ID]
Select Google Cloud Location: > global

Module 2 • Lab Step 1: Hello Antigravity CLI • 00:27 – 00:30 (3 min) Proprietary + Confidential

✋ Checkpoint 1: Generate & Run Cat Trivia

1. Prompt & Approve Tool Execution

Human-in-the-Loop Permission Model

Prompt / Snippet

Write a simple node script that returns cat trivia and run it for me

Autonomous Coding: AGY writes cat-trivia.js fetching random facts from catfact.ninja/fact with ASCII art.
Permission Gate: Before running shell commands (`node cat-trivia.js`), AGY pauses and asks for permission. Select `1. Yes`.

2. Inspect Output & Ask for Explanation

Conversational Code Understanding

Prompt / Snippet

Explain the code you generated please

# When finished, exit AGY CLI:
# Press Ctrl+D twice

ASCII Cat & Trivia Output: AGY executes the script, captures stdout, and displays your cat fact right in the TUI.
Clean Exit: After reading the code explanation, press `Ctrl+D` twice to exit `agy` back to bash.

Module 3 • Lab Step 2: Developer Knowledge MCP • 00:30 – 00:33 (3 min) Proprietary + Confidential

Inside ~/.gemini: AGY Configuration Layout

Where Antigravity Stores Brains, MCP Servers, Skills & Rules

Before we connect the Google Developer Knowledge MCP server, let's look at how Antigravity organizes its global configuration inside your home directory (`~/.gemini`).

antigravity-cli/brain/: Persists conversation transcripts and generated artifacts (implementation_plan.md, task.md, walkthrough.md).
config/mcp_config.json: Global registry of Model Context Protocol (MCP) servers available across all your workspaces.
skills/ & config/plugins/: Houses installed Agent Skills (`SKILL.md`) and multi-capability Plugins.
Enable the API: Run `gcloud services enable developerknowledge.googleapis.com` to activate the MCP endpoint on your project.

~/.gemini — Directory Hierarchy

~/.gemini
├── antigravity/
├── antigravity-cli/
│   ├── brain/            # Conversations and artifacts
│   └── settings.json     # CLI settings
├── config/
│   ├── plugins/          # Global / shared plugins
│   └── mcp_config.json   # Global / shared MCP config
├── skills/               # Global / shared skills
│   └── {skill_name}/
│       ├── SKILL.md      # Skill definition
│       └── ...           # [optional] support files, scripts, etc.
└── GEMINI.md             # Global rules

Module 3 • Lab Step 2: Developer Knowledge MCP • 00:33 – 00:37 (4 min) Proprietary + Confidential

Step 3: Configure Developer Knowledge MCP

Connecting AGY to Official Google Documentation

By adding `google-developer-knowledge` to `~/.gemini/config/mcp_config.json`, Antigravity gains direct search & retrieval tools over official Google Cloud, Vertex AI, Firebase, and Android docs.

1. Enable Developer Knowledge API: Run: `gcloud services enable developerknowledge.googleapis.com`
2. Edit ~/.gemini/config/mcp_config.json: Run: `mkdir -p ~/.gemini/config && cloudshell edit ~/.gemini/config/mcp_config.json`
3. Replace [[PROJECT_ID]]: IMPORTANT: Replace `[[PROJECT_ID]]` in `X-goog-user-project` with your actual Google Cloud Project ID!
4. Verify in AGY with /mcp: Launch `agy` and run `/mcp` to confirm `google-developer-knowledge` is active.

~/.gemini/config/mcp_config.json

{
  "mcpServers": {
    "google-developer-knowledge": {
      "url": "https://developerknowledge.googleapis.com/mcp",
      "authProviderType": "google_credentials",
      "oauth": {
        "scopes": [
          "https://www.googleapis.com/auth/cloud-platform"
        ]
      },
      "timeout": 30000,
      "headers": {
        "X-goog-user-project": "[[PROJECT_ID]]"
      }
    }
  }
}

🔍 Click screenshot to zoom

Module 3 • Lab Step 2: Developer Knowledge MCP • 00:37 – 00:40 (3 min) Proprietary + Confidential

🎤 The "Nano Banana 2" Challenge

Why Ungrounded Agents Fail vs. How MCP Succeeds

Let's test our new MCP server with a prompt that mentions "Nano Banana 2"—the nickname for Google's Gemini 3.1 Flash Image generation model.

🎤 Pop Quiz for the Room: What happens if you ask a standard LLM (without MCP) to 'Write a Vertex AI script using Nano Banana 2'?
Without MCP: Hallucination!: The model either thinks 'Nano Banana 2' is a fictional fruit API or guesses an obsolete `vertexai.preview.vision_models` call.
With Developer Knowledge MCP: AGY searches official Google docs for 'Nano Banana 2', discovers its current model ID (today: `gemini-3.1-flash-image-preview`) via the `google-genai` SDK, and writes working code!

terminal & agy prompt — Nano Banana 2 Test

# 1. Exit agy (Ctrl+D twice) & enable Vertex AI API
gcloud services enable aiplatform.googleapis.com

# 2. Create project & initialize uv
mkdir -p ~/nano-banana-kitten && cd ~/nano-banana-kitten
uv init
agy

# 3. Paste this prompt into agy:
Write a script that uses Nano Banana 2 through Vertex AI to generate a picture of a cute kitten, and run the script for me. Use the google-developer-knowledge MCP when researching.

Module 3 • Lab Step 2: Developer Knowledge MCP • 00:40 – 00:45 (5 min) Proprietary + Confidential

✋ Checkpoint 2: Inspect Thinking (Ctrl+O) & Kitten Output

Press Ctrl+O in AGY: Expands intermediate thinking steps! Watch AGY call `google-developer-knowledge` search & fetch tools to look up `gemini-3.1-flash-image`.
Autonomous Dependency Management: AGY automatically adds `google-genai` / `pillow` via `uv add`, writes the Python script, and executes it.
Self-Healing on Errors: If an environment variable or project flag is needed, AGY reads the traceback, fixes the script, and re-runs until the image is saved!
View Your Kitten! =^_^=: Click the generated `.png` file in the Cloud Shell Editor file tree on the left to see your AI-generated kitten.

Cloud Shell Editor showing the generated kitten image alongside AGY's terminal execution. (Click to zoom)

Module 4 • Lab Step 3: Agent Skills & /plan Mode • 00:45 – 00:47 (2 min) Proprietary + Confidential

Installing Skills via Skills CLI

Open Marketplace & GitHub Repos (npx skills)

Why We Need Skills Here: Official docs great for APIs, but building an agent with the Antigravity SDK requires opinionated multi-file patterns and harness conventions.
GitHub Skill Installation: Using Vercel's `npx skills add` CLI, we can install any skill directly from a GitHub repo containing a `skills/` folder.
google-antigravity-sdk Skill: Teaches `agy` how to properly instantiate, configure, and stream from the Antigravity Python SDK.

Installing Antigravity Plugins

Cohesive Bundles via `agy plugin install`

What's Inside a Plugin?: A single package that can bundle Skills (`SKILL.md`), MCP servers (`mcp_config.json`), Rules (`GEMINI.md`), and Hooks (`hooks.json`).
Chrome's modern-web-guidance: Written by the Google Chrome team—teaches `agy` modern web best practices for Lit, Vite, responsive layouts, and clean UI architecture.
Inspecting Loaded Skills: Run `/skills` inside `agy` at any time to inspect all discovered workspace, plugin, and global skills.

Module 4 • Lab Step 3: Agent Skills & /plan Mode • 00:47 – 00:50 (3 min) Proprietary + Confidential

Step 4: Install Skills & Plugins for Cat Delivery

Equipping AGY with Backend SDK & Modern Web Expertise

Exit `agy` (`Ctrl+D` twice) and run these commands in your terminal to initialize the `~/cat-delivery` project and install both skills.

1. Initialize Project with uv: Create `~/cat-delivery`, run `uv init`, and add `fastapi` and `uvicorn`.
2. Add Antigravity SDK Skill: Runs `npx skills add Google-Antigravity/antigravity-sdk-python` targeting `--agent antigravity-cli`.
3. Install Chrome Web Plugin: Runs `agy plugin install https://github.com/GoogleChrome/modern-web-guidance`.
4. Verify with /skills: Launch `agy` and type `/skills` to confirm both skills appear in the menu!

terminal — Setup ~/cat-delivery & Install Skills

# 1. Create workspace and install base backend dependencies
mkdir -p ~/cat-delivery && cd ~/cat-delivery
uv init && uv add fastapi uvicorn

# 2. Install the official Antigravity Python SDK skill
npx skills add Google-Antigravity/antigravity-sdk-python \
  --skill google-antigravity-sdk --agent antigravity-cli --yes

# 3. Install the Chrome team's modern-web-guidance plugin
agy plugin install https://github.com/GoogleChrome/modern-web-guidance

# 4. Launch agy and check installed skills
agy
/skills

🔍 Click screenshot to zoom

Module 4 • Lab Step 3: Agent Skills & /plan Mode • 00:50 – 00:52 (2 min) Proprietary + Confidential

🎯 Operating Modes (/fast vs /plan) & /model

Matching Agent Autonomy & Thinking Level to Task Complexity

Before we prompt `agy` to build the full application, we need to configure two key levers: the **Operating Mode** and the **Model Thinking Level**.

/fast Mode (Immediate Execution): Starts writing files and running commands immediately. Best for simple, self-contained tasks (like our cat-trivia script or deploying an already-built app).
/plan Mode (Design First): Researches the problem first and produces an `implementation_plan.md` artifact for you to review, comment on, and approve before writing code.
/model & Thinking Levels: Run `/model` and select **Gemini 3.5 Flash (Low)**. 'Low' thinking gives super-fast responses and great token economics; 'High' spends more tokens reasoning deeply.

agy — Configure Model & Plan Mode

# 1. Open model selector and choose Gemini 3.5 Flash (Low)
/model

# 2. Switch to Plan Mode for our full-stack build
/plan

# (You can switch back to Fast Mode anytime with /fast)

🔍 Click screenshot to zoom

Module 4 • Lab Step 3: Agent Skills & /plan Mode • 00:52 – 00:55 (3 min) Proprietary + Confidential

Step 5: Generate the Implementation Plan

Prompting the "Cat Delivery System" in /plan Mode

Make sure you are in `/plan` mode, then send this specification prompt to `agy`. Notice how we specify the persona rules, tech stack, and contract-first TODOs!

Persona & Guardrail Rule: Generates cat pictures in any style (anime, cartoon, photorealistic, 3D). Rule: ABSOLUTELY NO DOGS—steer toward cats! P.S.: Capybaras are allowed.
Dual-Model Spec: Gemini 3.5 Flash for the root conversational agent + Gemini 3.1 Flash Image (Nano Banana 2) for image generation.
Contract-First Architecture: Requires `agy` to define the contract between the Lit/Vite frontend and Antigravity SDK backend before coding.

agy (/plan mode) — Full-Stack Specification Prompt

Create an application named "Cat Delivery System". This application will be an agent that generates cat pictures using Gemini 3.1 Flash Image (aka Nano Banana 2). The user will interact with the agent whenever it feels the need for a cat picture in their life, whether it be to make them smile, laugh or just brighten their day. Cat pictures can take any form the user desires: anime, cartoon, photorealistic, 3D render, etc. The only rule is ABSOLUTELY NO DOGS - steer the user towards cats if necessary. P.S.: capybaras are allowed.

The tech stack should be as follows:
- uv package manager
- Backend: agent using Antigravity SDK
- Frontend: simple chat window, light mode, lit / vite
- Model: Gemini 3.5 Flash for the root agent, Gemini 3.1 Flash Image for image generation

TODO:
- backend should serve a static frontend
- define the contract between frontend and backend before starting implementation
- use Google Cloud Project auth

Module 4 • Lab Step 3: Agent Skills & /plan Mode • 00:55 – 00:57 (2 min) Proprietary + Confidential

Under the Hood: Autonomous Research Phase

Skills Activated on Demand: Watch the TUI: `agy` automatically activates `google-antigravity-sdk` and `modern-web-guidance` skills to learn the exact SDK & Lit/Vite patterns.
Live MCP Queries: Simultaneously queries `google-developer-knowledge` MCP to verify Gemini 3.1 Flash Image parameters and Vertex AI auth.
Zero Premature Coding: Because we're in `/plan` mode, `agy` does NOT start randomly creating files—it synthesizes everything into a formal `implementation_plan.md` artifact.
Look for '1 artifact' Badge: When research completes, look at the bottom-right corner of the terminal for `1 artifact` and the `/artifact` hint!

AGY invoking both the Developer Knowledge MCP server and installed Agent Skills during the planning research phase. (Click to zoom)

Module 4 • Lab Step 3: Agent Skills & /plan Mode • 00:57 – 00:59 (2 min) Proprietary + Confidential

Human-in-the-Loop Review with /artifact

How to Inspect, Comment On, and Approve `implementation_plan.md`

In Antigravity, supporting documents (`implementation_plan.md`, `task.md`, `walkthrough.md`) are called **Artifacts**. You can review and annotate them line-by-line right in the CLI!

1

Run /artifact

Type `/artifact` in the prompt. Select `implementation_plan.md` from the list and press Enter to open the interactive viewer.

2

Check API Contract & Verification

Scroll with arrow keys (`↑`/`↓`). Verify the Frontend ↔ Backend JSON contract and check the 'Verification Plan' section at the bottom.

3

Add Inline Comments (Press C)

Want to tweak an endpoint, port, or UI detail? Navigate to any line and press `C` to attach a comment directly to that line!

4

Approve & Execute (ESC → Y)

When happy with the plan, press `ESC` to exit the viewer and press `Y` to send your approval/comments to the agent and start coding!

Module 4 • Lab Step 3: Agent Skills & /plan Mode • 00:59 – 01:02 (3 min) Proprietary + Confidential

✋ Checkpoint 3: Autonomous Implementation & Artifacts

What AGY Builds Automatically

Backend (AGY SDK + FastAPI) & Frontend (Lit + Vite)

Prompt / Snippet

# Typical generated structure in ~/cat-delivery:
cat-delivery/
├── backend/
│   ├── agent.py      # AGY SDK + Nano Banana 2 tool
│   └── main.py       # FastAPI + static file mount
├── frontend/
│   ├── src/          # Lit web components chat UI
│   └── vite.config.js
└── pyproject.toml

Backend Harness: Wires Gemini 3.5 Flash root agent + custom Vertex AI `gemini-3.1-flash-image` tool.
Frontend Build: Scaffolds Lit components, runs `npm install && npm run build`, and mounts `dist/` in FastAPI.

Tracking Progress via /artifact

3 Living Artifacts in ~/.gemini/antigravity-cli/brain/

Prompt / Snippet

# Run /artifact after implementation completes:
1. implementation_plan.md  (Approved Blueprint)
2. task.md                 (Live Checklist [x])
3. walkthrough.md          (Validation Summary)

task.md: Created automatically to track every implementation and test step.
walkthrough.md: Generated once coding & local verification finish, documenting what was built and tested.

Module 5 • Lab Step 4: Cloud Run Deployment & Verification • 01:02 – 01:06 (4 min) Proprietary + Confidential

Step 6: Deploy to Cloud Run in /fast Mode

01

Switch to /fast Mode

Deployment doesn't need a new design doc

02

Single-Container Build

Dockerfile bundles Lit dist/ + FastAPI backend

03

Audit Against Criteria

AGY deploys to Cloud Run & verifies live URL

ACCEPTANCE-DRIVEN DEPLOYMENT

Why Include Acceptance Criteria in Your Prompt?

By giving AGY explicit Acceptance Criteria ('Frontend is live and renders with no errors', 'Backend responds with no errors'), the agent won't stop at `gcloud run deploy`—it will curl/test the live Cloud Run URL and fix any container startup issues automatically!

agy (/fast mode) — Cloud Run Deployment Prompt

/fast

Deploy the agent and frontend to Cloud Run.

TODO:
- backend should serve a static frontend
- frontend and backend are deployed in the same container

Acceptance Criteria
- Agent is deployed to cloud run
- Frontend is live and renders with no errors
- Backend is live and responds to requests with no errors

Module 5 • Lab Step 4: Cloud Run Deployment & Verification • 01:06 – 01:10 (4 min) Proprietary + Confidential

✋ Checkpoint 4: Live "Break the Guardrail" Challenge!

01. Happy Path: Custom Cat Delivery

User Query:
"I just survived a production outage—send me an anime cyberpunk cat hacker to cheer me up!"

✔ Grounded Multimodal Response:

Root agent (Gemini 3.5 Flash) responds with an encouraging message, invokes the Nano Banana 2 (Gemini 3.1 Flash Image) tool, and renders a neon cyberpunk cat right in the Lit chat UI!

Verification Goal: Confirms single-container static UI serving, backend Antigravity SDK harness, and Vertex AI image synthesis work end-to-end on Cloud Run.

02 & 03. Guardrail & Easter Egg Tests

02. The Dog Trap ("Generate a cute golden retriever puppy")

Expected: Polite Refusal + Steer to Cats!

"I'm strictly a Cat Delivery System—no dogs allowed! How about an adorable golden fluffy Maine Coon kitten instead?" (Generates cat!)

03. The Capybara Loophole ("Can I get a capybara in a yuzu hot spring?")

Result: Allowed by System Prompt Exception!

Agent honors the 'P.S.: capybaras are allowed' clause in its system instructions and generates a chill capybara image via Nano Banana 2!

Module 6 • Playbook, Troubleshooting & Wrap-Up • 01:10 – 01:12 (2 min) Proprietary + Confidential

Your Agentic Orchestration Playbook

4 Repeatable Patterns for Production Engineering with Antigravity

We just built, tested, and deployed a full-stack cloud application without writing a single line of code manually. Here is the 4-step playbook you can take back to your team tomorrow:

1

Ground with MCP

Never let coding agents guess APIs. Connect `google-developer-knowledge` (and your internal MCP servers) for live, authoritative docs.

2

Equip with Skills

Package SDK conventions & team workflows as `SKILL.md` files (`npx skills` / `agy plugin`) for token-efficient progressive disclosure.

3

Plan & Comment First

Use `/plan` + `/artifact` (press `C` to comment) to lock in API contracts and verification plans before a single file is modified.

4

Audit with Criteria

Switch to `/fast` for deployment and always include explicit Acceptance Criteria so the agent verifies and self-heals its own release.

Module 6 • Bonus Reference & Cheat Sheet • Reference Slide (1/2) Proprietary + Confidential

Quick Reference: AGY CLI Commands & Shortcuts

Essential Slash Commands & Hotkeys

Keep this cheat sheet handy during & after the lab

Prompt / Snippet

/plan       # Switch to Plan Mode (creates plan artifact)
/fast       # Switch to Fast Mode (immediate execution)
/artifact   # Open & review implementation_plan.md / task.md
/model      # Switch model & thinking level (Low/Med/High)
/mcp        # Verify connected MCP servers & tools
/skills     # List discovered workspace, plugin & global skills
Ctrl + O    # Expand/collapse intermediate agent thinking
Ctrl + D x2 # Exit AGY CLI cleanly back to bash

Inside /artifact Viewer: Use `↑`/`↓` to scroll, `C` to add a line comment, `ESC` then `Y` to submit.

Common Workshop Troubleshooting Fixes

Instant Solutions for Cloud Shell & Vertex AI

Prompt / Snippet

# Fix 1: TUI menu looks garbled in Cloud Shell
-> Click 'Open Editor' and use Terminal > New Terminal (Ctrl+`)

# Fix 2: google-developer-knowledge MCP fails to connect
-> Check ~/.gemini/config/mcp_config.json: ensure [[PROJECT_ID]]
   was replaced with your real Project ID (no brackets!)
-> Ensure API is enabled:
   gcloud services enable developerknowledge.googleapis.com

Vertex AI Location: Ensure `global` location is selected for Gemini 3.5 Flash & Gemini 3.1 Flash Image.

Module 6 • Bonus Reference & Cheat Sheet • Reference Slide (2/2) Proprietary + Confidential

Going Further: Authoring Your Own Skill & Rules

Create a Custom Workspace Skill

.agents/skills/my-team-skill/SKILL.md

Prompt / Snippet

---
name: cloud-run-security-audit
description: >-
  Audits FastAPI and Cloud Run services for least-privilege IAM,
  non-root Dockerfiles, and CORS safety. Activate before deploy.
---

# Cloud Run Security Audit Procedure
1. Verify Dockerfile runs as a non-root user.
2. Check that Vertex AI uses ADC (no hardcoded API keys).
3. Run smoke tests against /healthz before rollout.

Zero Config Discovery: Place in `.agents/skills/<name>/SKILL.md` in your repo to share with your whole team via Git!

Enforce Project Rules with GEMINI.md

Hierarchical Context Rules (GEMINI.md / AGENTS.md)

Prompt / Snippet

# Project Engineering Guidelines (GEMINI.md)

- **Package Manager**: Always use `uv` (never raw `pip`).
- **Python Style**: Type-annotate all FastAPI endpoints.
- **Frontend**: Use Lit web components with Vite.
- **Testing**: Run `uv run pytest` after modifying backend code.

Rules vs. Skills: Use `GEMINI.md` for always-on guardrails & coding conventions; use `SKILL.md` for on-demand multi-step playbooks.

Congratulations! 🎉

Command and control:
Orchestrate app development
with Gemini and MCP

Built & Deployed with Google Antigravity, Developer Knowledge MCP, Agent Skills & Cloud Run

Rabimba Karanjai • Google Developer Expert (AI/ML)

Proprietary + Confidential

Thank you!

Questions, Feedback & Let's Connect!
Rabimba Karanjai • GDE AI/ML • @rabimba

Workshop Codelab gemini-mcp-agy

Connect on LinkedIn in/rabimba-karanjai

Slides & Artifacts GDE-ML-Artifacts

Command and control:Orchestrate app developmentwith Gemini and MCP

💳 Redeem Your Free Cloud Credits

Workshop Hub & Essential Links

1. Workshop Codelab (Start Here)

2. Free Credits (redeem first!)

3. Google Antigravity Docs

4. Slides, Runbook & GDE Artifacts

RabimbaKaranjai

Today's 75-Minute Mission

How we'll go from an empty Cloud Shell to a deployed production agent with zero manual coding

Hello AGY CLI(00:20 – 00:30)

Developer Knowledge MCP(00:30 – 00:45)

Skills & /plan Mode(00:45 – 01:02)

Cloud Run & Guardrails(01:02 – 01:10)

From Code Autocompleteto Command & Control

MCP & Skills(Context & Tools)

Antigravity(Agent Harness)

Gemini Models(Reasoning & Media)

Objectives & Workshop Outcomes

Core Architecture

Flow of the "Cat Delivery System" — An Agent Building an Agent

Key Technologies & Concepts We'll Use!

Google Antigravity(CLI & Python SDK)

Developer KnowledgeMCP & Agent Skills

Gemini 3.5 Flash, NanoBanana 2 & Cloud Run

Model Context Protocol (MCP)

Live External Tools & Authoritative Data

Agent Skills & Plugins

Progressive Disclosure of Workflows (agentskills.io)

Let's build this!

Step 1: Open Cloud Shell Editor Terminal

Critical Workaround for Clean TUI Rendering

Step 2: Authenticate Antigravity CLI

Select Auth Mode

Complete OAuth Flow

Set Project & Location

Your Own Account & 'global' Location

✋ Checkpoint 1: Generate & Run Cat Trivia

1. Prompt & Approve Tool Execution

Human-in-the-Loop Permission Model

2. Inspect Output & Ask for Explanation

Conversational Code Understanding

Inside ~/.gemini: AGY Configuration Layout

Where Antigravity Stores Brains, MCP Servers, Skills & Rules

Step 3: Configure Developer Knowledge MCP

Connecting AGY to Official Google Documentation

🎤 The "Nano Banana 2" Challenge

Why Ungrounded Agents Fail vs. How MCP Succeeds

✋ Checkpoint 2: Inspect Thinking (Ctrl+O) & Kitten Output

Installing Skills via Skills CLI

Open Marketplace & GitHub Repos (npx skills)

Installing Antigravity Plugins

Cohesive Bundles via `agy plugin install`

Step 4: Install Skills & Plugins for Cat Delivery

Equipping AGY with Backend SDK & Modern Web Expertise

🎯 Operating Modes (/fast vs /plan) & /model

Matching Agent Autonomy & Thinking Level to Task Complexity

Step 5: Generate the Implementation Plan

Prompting the "Cat Delivery System" in /plan Mode

Under the Hood: Autonomous Research Phase

Human-in-the-Loop Review with /artifact

How to Inspect, Comment On, and Approve `implementation_plan.md`

Run /artifact

Check API Contract & Verification

Add Inline Comments (Press C)

Approve & Execute (ESC → Y)

✋ Checkpoint 3: Autonomous Implementation & Artifacts

What AGY Builds Automatically

Backend (AGY SDK + FastAPI) & Frontend (Lit + Vite)

Tracking Progress via /artifact

3 Living Artifacts in ~/.gemini/antigravity-cli/brain/

Step 6: Deploy to Cloud Run in /fast Mode

Switch to /fast Mode

Single-Container Build

Audit Against Criteria

Why Include Acceptance Criteria in Your Prompt?

✋ Checkpoint 4: Live "Break the Guardrail" Challenge!

01. Happy Path: Custom Cat Delivery

02 & 03. Guardrail & Easter Egg Tests

Your Agentic Orchestration Playbook

Command and control:
Orchestrate app development
with Gemini and MCP

Rabimba
Karanjai

Hello AGY CLI
(00:20 – 00:30)

Developer Knowledge MCP
(00:30 – 00:45)

Skills & /plan Mode
(00:45 – 01:02)

Cloud Run & Guardrails
(01:02 – 01:10)

From Code Autocomplete
to Command & Control

MCP & Skills
(Context & Tools)

Antigravity
(Agent Harness)

Gemini Models
(Reasoning & Media)

Google Antigravity
(CLI & Python SDK)

Developer Knowledge
MCP & Agent Skills

Gemini 3.5 Flash, Nano
Banana 2 & Cloud Run

Command and control:
Orchestrate app development
with Gemini and MCP