| name | ge-demo-generator |
| description | Synthesizes and deploys complete, domain-specific Gemini Enterprise demo environments directly to Google Cloud. Use when the user asks to create an AI agent demo for any customer domain (e.g. 'example.com', 'example.co.jp', 'example.de', 'example.fr' - any company, any industry, any region) or business goal, generate realistic BigQuery/Firestore sample datasets, create external demo files (PDF, Excel, scanned images), stage them in Cloud Storage and upload them to the deploying account's Google Drive, scaffold ADK multi-agent architectures with MCP tools and A2UI cards, deploy to Cloud Run, publish to Gemini Enterprise, and generate 7 structured demo prompts in any language. Confirms the requirements interactively and presents a demo architecture & data model plan (Mermaid ER diagram, external file lineage, target project) for approval before anything is deployed. Also triggered by /ge-demo-generator. |
| metadata | {"author":"Google Cloud Customer Engineering","version":"2.15.0"} |
GE Demo Generator Skill (v2.15.0)
Synthesizes production-grade, domain-tailored AI agent demo environments using Gemini 3.8 Flash for reasoning and Gemini 3.1 Flash Image for visual generation, adhering to a strict 6-step infrastructure dependency graph, rich A2UI interactive component streaming, Google Workspace OAuth authorization, external sample files staged in Cloud Storage and, when the credentials carry the Drive scope, in the deploying account's Google Drive, 7 structured demo prompts, and global multilingual localization (i18n/l10n).
🌐 Multilingual Support & Localization Architecture (i18n/l10n)
The skill is implemented entirely in English at the system specification and codebase layer, while providing complete, dynamic multilingual support for any enterprise domain worldwide:
-
Automatic Language & Locale Detection:
- From Domain TLD:
.co.jp, .jp, .ne.jp, .or.jp ➔ Japanese (日本語, locale: "ja", currency: "JPY", symbol: "¥")
.de ➔ German (Deutsch, locale: "de", currency: "EUR", symbol: "€")
.fr ➔ French (Français, locale: "fr", currency: "EUR", symbol: "€")
.es ➔ Spanish (Español, locale: "es", currency: "EUR", symbol: "€")
.it ➔ Italian (Italiano, locale: "it", currency: "EUR", symbol: "€")
.cn, .tw ➔ Chinese (中文, locale: "zh", currency: "CNY"/"TWD")
.kr ➔ Korean (한국어, locale: "ko", currency: "KRW", symbol: "₩")
.br ➔ Portuguese (Português, locale: "pt", currency: "BRL", symbol: "R$")
.co.uk, .com.au, .com, .io, .org, etc. ➔ English (locale: "en", currency: "USD"/"GBP")
- From User Interaction:
- If the user provides a prompt in Japanese, all user-facing demo assets are synthesized in Japanese. If in German, all in German.
-
Strict Language Consistency Rule (MANDATORY):
- ALL user-facing text elements MUST strictly match the detected target language:
Complete Lifecycle & Infrastructure Dependency Graph
Phase 1: Customer Domain & Business Goal Research (Search Grounding & Language Detection)
↓
Phase 2: Requirements confirmed interactively ➔ Demo Architecture & Data Model Plan ➔ APPROVAL
↓
Phase 3: Synthetic Data & External Sample Files Generation (PDF, Excel, Images) + Google Drive Upload (Step 1/6)
↓
Phase 4: ADK Multi-Agent Project Scaffolding & A2UI System Prompts
↓
Phase 5: Ordered Cloud Provisioning & Deployment:
├── Step 1: BigQuery (with Knowledge Catalog Metadata) & Firestore Initial Data
├── Step 2: Agent Engine Sandbox (Code Execution Environment) [CRITICAL DEPENDENCY]
├── Step 3: Data Viewer Dashboard Cloud Run Deployment (Gets VIEWER_URL)
├── Step 4: Main Multi-Agent Cloud Run Deployment (Injects SANDBOX_RESOURCE_NAME & VIEWER_URL)
└── Step 5: Background Task Pub/Sub Push Subscription (/execute_task & SELF_URL)
↓
Phase 6: Gemini Enterprise App Discovery, Workspace Authorization & Registration (Step 6/6)
↓
Phase 7: Comprehensive Results Output:
├── 💬 Direct Gemini Enterprise Console Chat Link
├── 📁 External Sample Files: Cloud Storage links, plus Drive links when the upload ran
├── 📊 Firestore Data Viewer Dashboard Link
├── 🔎 BigQuery Console Link
└── 🎯 7 Structured Demo Prompts Playbook (Localized to Target Language)
Phase 1: Customer Domain & Business Goal Research
-
Extract Domain & Company:
- If a domain is provided (e.g.,
example.com, example.co.jp, example.de), determine company name, primary industry, and regional language from the domain itself — never from a list baked into this skill.
- If a business goal or company name is provided, identify the corresponding corporate entity and language.
-
Grounded Deep Research (via search_web):
- Search for the company's latest corporate profile, business units, IR reports, and known operational challenges.
- Identify:
- 3-5 Key Business Challenges (e.g., "yield rate drop in semiconductor fabrication", "supply chain lead-time anomalies").
- 5-8 Operational Workflows and evaluate which ones are automatable with AI agents.
- Authentic Domain Entities & Nomenclature (e.g., real facility names like "Kumamoto Technology Center", actual product codes, industry status codes).
-
Interactive Use Case Selection (Customer Domain Selection Flow):
- Present the research findings to the user in their language:
- 🏢 Company Profile: Official name, summary, primary industry.
- ⚠️ Key Business Challenges: 3-5 detected operational pain points.
- 🎯 Candidate Workflows: 3-5 top automatable workflows with automation feasibility ratings.
- Prompt the user to select or customize the use case:
- Use
ask_question or present a structured choice so the user can pick the target scenario or write in specific requirements.
Phase 2: Requirements Alignment & Demo Architecture Plan (Approval Gate)
Phase 2 has two steps, in this order and never merged into one:
Step 2.1 — ask the handful of questions the design genuinely depends on, interactively.
Step 2.2 — show the finished design brief as the alignment artifact, and ask for a go.
The order is what makes the brief worth reading. A brief that still contains open questions
is a questionnaire, and the user has to hold the design in their head while answering it; a
brief written after the answers is a mirror — the user is checking whether what you
understood matches what they meant, which is a much easier thing to do and the last cheap
moment to change the entities, the narrative, the file lineage or the target project.
Until the user approves the brief, nothing is created — no CSVs, no Drive folder, no
BigQuery dataset, no Cloud Run service. Everything after this point takes 15-30 minutes and
spends real quota in someone's project.
Step 2.1 — Confirm the requirements interactively
Ask only what you cannot responsibly decide, and ask it before writing the brief. Keep it
short — two to four questions, batched into one message (ask_question where available), each
with a stated default so the user can answer "all defaults" in three words:
- Anchor persona — who is the primary user of this agent (job title, and a sentence of
context if useful)? That persona becomes the demo's protagonist: the agent is framed as
their process operator, prompt 1 opens from their daily situation, approval gates are
decisions they or their manager own, and every other persona appears only as a hand-off
counterpart. If the user declines, rotate personas freely instead.
- Scope of the narrative — which single process instance the demo follows, if Phase 1
left more than one plausible reading of the chosen scenario.
- Advanced settings — ask about the two or three the demo is a plausible candidate for
(Workspace OAuth when the scenario touches Gmail/Drive/Calendar,
rag when it turns on
documents rather than numbers, dataScale when the narrative claims enterprise volume),
and say the rest keep their defaults. Keep the question short — brief section 6 lists
every option with its resolved value, so nothing is hidden by asking about a few. Add
a warm Cloud Run instance (MIN_INSTANCES=1) to that question whenever the user has
named a date, an audience or a live session: it is the one option that is about the
presentation rather than the demo, and the only thing that removes the cold start — and
the occasional cold-start error — from the first message. Offer it with its price in the
same breath (see brief section 6).
- Target environment — only if what
gcloud reports (read it now, see below) is not
obviously the project the user means.
Anything the user has already stated — in the original request or in Phase 1 — is decided.
Re-asking it reads as not having listened.
Read the target environment before writing the brief, so its section 5 states facts rather
than intentions:
GCP_ACCOUNT=$(gcloud config get-value account 2>/dev/null || echo "Unknown")
PROJECT_ID=$(gcloud config get-value project)
PROJECT_NUMBER=$(gcloud projects describe "$PROJECT_ID" --format="value(projectNumber)")
REGION=${CLOUD_RUN_REGION:-"asia-northeast1"}
# Whether the sample documents can reach a Google Drive at all - sections 3 and 5
# depend on it. The deploy uploads them with this same token, so this one call
# answers the question exactly as the deploy will.
DRIVE_OK=$(curl -s -o /dev/null -w '%{http_code}' \
-H "Authorization: Bearer $(gcloud auth print-access-token 2>/dev/null)" \
'https://www.googleapis.com/drive/v3/about?fields=user')
DRIVE_OK is 200 when the Drive copy will happen, and the owner is ${GCP_ACCOUNT} — the
same account, because the upload uses this token. Anything else (usually 403, a token
carrying no Drive scope, which is what a plain gcloud auth login gives you) means this
demo will have no Google Drive copy of the sample documents at all. That is a fact the
brief has to state, not discover at deploy time — see brief sections 3 and 5. It is also one
command to fix, so say it now: gcloud auth login --enable-gdrive-access --no-launch-browser (using --no-launch-browser since an agentic IDE terminal usually has no local browser), then re-read.
Step 2.2 — Present the Demo Architecture & Data Model Plan
One message, six sections, in this order, then the gate. Every answer from Step 2.1 is already
folded in — the brief states decisions, it does not ask for them, and it contains no
"tell me which…" or "TBD" anywhere.
Write it in the demo's language (§ i18n above) — headings, table headers and prose alike.
Only the technical identifiers stay English snake_case: table names, column names and
.env keys.
Open with a title line naming the company and the selected scenario, so the brief is
self-contained when it is pasted into a chat with a colleague:
🏗️ <Company> — Demo Architecture & Data Model Plan
Based on the selected scenario "<scenario>", here is the data layer design, the
Mermaid ER diagram and the cross-source reconciliation lineage.
§ Brief section 1 — 🏢 Target Company & Demo Overview
- Company / group: official name, and the group entity the demo actually models.
- Industry: the specific one, in the company's own vocabulary.
- Demo narrative: 2-3 sentences following one business process instance (one
problematic order, one flagged application) end to end, naming the departments it passes
through. All 7 demo prompts later trace this same instance; see
references/demo_prompts_guide.md §1.1-§1.2.
- Anchor persona: the persona confirmed in Step 2.1, stated as a fact ("the agent is
operated by the production control manager of the main works"). If the user declined to
name one, say that personas rotate instead — do not re-ask here.
§ Brief section 2 — 📊 Data Layer Design & Mermaid ER Diagram
- Render a Mermaid
erDiagram in a ```mermaid fenced block: every table, every column
with its type, PK/FK markers, cardinality labels on each relationship, and a short
business gloss in quotes after each column (string order_id PK "order number (ORD-2026-XXXX)").
The gloss is the same sentence that later becomes the column description in the Knowledge
Catalog, so writing it here is not duplicated work.
- Include the Firestore task collection as an entity too — it is part of the model the demo
reasons over, even though it is not in BigQuery.
- Shared core-system model: design the tables as a slice of the company's system of
record, not one team's dataset — a transactional table two departments both write, a
rules/threshold table owned by one department that another's workflow must consult, and
an audit seed visible only when the departments' views are joined. Details and the
single-department escape hatch:
references/data_generation.md §1.3.
§ Brief section 3 — 📁 External Sample Files & Cross-Source Lineage
Title this section after where the files will actually be, not after where people assume they
go. Cloud Storage is the only guaranteed destination (below), so a heading that says "Google
Drive" is wrong in every run whose token lacks the Drive scope — and it was, in most of them.
A table with one row per external file — type, filename, and what binds it to BigQuery:
| External file | Filename | Content & data binding |
|---|
| 📄 PDF audit report | <domain>_audit_report.pdf | which tables it summarizes, and the deliberate 5-15% variance that makes cross-source reconciliation necessary |
| 📊 Excel ledger | <domain>_external_ledger.xlsx | which FK column it joins on, and the row count |
| 🖼️ Scanned form 1 | handwritten_order_1.jpg | what the form is, in-domain, generated by gemini-3.1-flash-image |
| 🖼️ Scanned form 2 | handwritten_order_2.jpg | the second form, and the discrepancy it carries |
State the join key explicitly. A lineage row that says "related to orders" is not a design —
the whole point of the external files is that a question can only be answered by reading a
document and querying a table, and that only works if the keys line up (70%+ FK match,
5-15% audit variance). See references/external_files_and_drive.md.
Close the section with where the files will actually live, in one line, because it is the
part users assume wrongly. Write the line that matches the DRIVE_OK read above:
DRIVE_OK | The line to write |
|---|
200 | Staged to gs://<project>-<domain>-<suffix>-docs/ and uploaded at deploy time to the Google Drive of ${GCP_ACCOUNT}, which owns the folder. |
| anything else | Staged to gs://<project>-<domain>-<suffix>-docs/. No Google Drive copy — this machine's gcloud credentials carry no Drive scope, so a Drive or Sheets step in the demo script would find nothing. |
The Cloud Storage staging happens in both mcp and rag mode; the deploy-time upload is
the only path into any Drive. The deployed agent has no way to put them in anyone's Drive —
v2.11.0 removed the in-conversation import, deliberately. If the user wants them in Drive and
DRIVE_OK is not 200, say so here, while it can still be sorted out
(gcloud auth login --enable-gdrive-access --no-launch-browser and re-run, or upload ./external_files/ by hand
afterwards).
§ Brief section 4 — 🤖 Agent Profile & Runtime
- Agent Name (
DEMO_DISPLAY_NAME): Concise 2–4 word domain role (e.g. TWG Tea Retail Operations Director, Mercari Trust & Safety Specialist). Registered in Gemini Enterprise as ${DEMO_DISPLAY_NAME} (${SERVICE_NAME}), matching the Web UI (GAS) version.
- Description (
DEMO_DESCRIPTION): Concrete, professional 1–2 sentence mission summary specifying the business domain, core datasets, and operational goals (e.g. Orchestrates boutique inventory balancing, central commissary replenishment, and plantation harvest orders across Singapore.). Matches oneSentenceSummary in the Web UI (GAS).
- Reasoning / orchestration:
gemini-3.8-flash for all agent instances (Root Coordinator, Deep Analysis Sub-Agent, Background Worker).
- Image generation:
gemini-3.1-flash-image (with localized prompt reinforcement and GCS artifact persistence).
- Code execution: Agent Engine Sandbox (
us-central1 — fixed, not $REGION).
- UI: Gemini Enterprise A2UI v0.9 composite catalog (name the components this demo will actually lean on —
MaterialTable, VegaChart, MaterialCard).
§ Brief section 5 — ⚙️ Deployment Target Environment
Show what the commands above returned, verbatim, in a fenced block:
👤 Active User Account : ${GCP_ACCOUNT}
🏢 Target Project : ${PROJECT_ID} (${PROJECT_NUMBER})
🌐 Target Region : ${REGION}
🗂️ Sample File Storage : gs://${PROJECT_ID}-<domain>-<suffix>-docs/
📁 Google Drive Copy : <the DRIVE_OK line, below>
This is the section users most often stop on, because the answer is frequently "wrong
project". Do not paraphrase the values and do not print a project the user named unless
gcloud agrees — the deploy will use what gcloud says, not what the brief claims.
Both storage lines are mandatory, and neither is optional wording: the Cloud Storage bucket is
created in every mode, and the Drive line is the one users read as a promise. Write it from
DRIVE_OK, never from intent:
DRIVE_OK | 📁 Google Drive Copy |
|---|
200 | Google Drive of ${GCP_ACCOUNT} |
| anything else | none - these credentials have no Drive scope (fix: gcloud auth login --enable-gdrive-access --no-launch-browser) |
Never print a Drive destination the deploy cannot reach, and never describe the copy as
something the agent will do later — nothing after the deploy writes to a Drive. See
references/external_files_and_drive.md §5.
§ Brief section 6 — 🎛️ Advanced Settings & Integrations
State the settings this demo will deploy with — the answers from Step 2.1, plus the
defaults for everything the user did not touch. These names are the .env keys,
upper-snake-cased: enableManagedAgent is ENABLE_MANAGED_AGENT. Every one of them is a
real switch in the deployed container, so an option discussed here but not written to
.env is a feature the demo will not have.
List all nine, every time, in this order, with the value this deploy will use in the
second column — ✅ true / ❌ false / the literal value / — (unset). An option the
brief leaves out is an option the user cannot ask for: they do not know it exists, and by
the time the deploy banner mentions it the 15-30 minutes are already spent. The gate is the
last cheap moment to turn one on, so the gate has to show the whole board. Keep the third
column as written here — it is what makes a false decidable rather than merely visible —
and reword only where this demo's scenario changes what an option would buy.
| Option | This demo | Default | What it buys, and what it costs |
|---|
🤖 enableManagedAgent | <value> | true | Agent Engine Sandbox code execution, asynchronous background delegation (delegate_autonomous_task), scheduled tasks and Drive deliverable exports. The delegation prompts in the demo playbook exercise this, which is why it is the one default-on capability. Adds ~8-10 min of provisioning, overlapped with the rest of the deploy. |
🔎 dataExplorationMode | <value> | mcp | How the agent reads the demo's data. mcp (default) provisions no search index: the data-asset catalog is already in the agent's system instruction, so a figure question is one execute_sql call — the four-to-five round trips people blame on "no index" came from the metadata expedition in front of the query, and the mcp routing block overrides exactly that. rag additionally builds a Discovery Engine index over the BigQuery dataset and the staged files and makes it the read path: lookups and document questions return in one sub-second search_datastore call, while computed figures, joins and every write stay on MCP because the index lags the tables. Pick rag when the demo turns on documents rather than on numbers, and note it also attaches data stores to the (often shared) Gemini Enterprise app — see references/datastore_connectors.md. |
📁 enableDatastoreFs | <value> | false | When dataExplorationMode=rag, provisions a semi-structured Discovery Engine DataStore (ds-${SERVICE_NAME}-fs) from FIRESTORE_COLLECTION via FirestoreSource (GCS export staging). Enables semantic search over historical incident tickets, resolved remediation logs, and SOP archives. Note: live task mutations, approvals, and Operations Viewer synchronization continue to use Firestore MCP for sub-100ms real-time state tracking. |
🔑 enableWorkspaceAuth | <value> | false | User-OAuth passthrough — the agent acts as the signed-in user for the Drive handoff and Workspace token plumbing. Commonly wanted, since Workspace is usually available in the target environment, but default-on: some organizations refuse to authorize an OAuth client they have not vetted, and there sign-in fails for every demo user. Confirm the target org permits it before enabling. |
Only list the options that are on, plus the two or three the demo is a plausible candidate
for. A user reading eight defaults they did not ask about is a user who skims the whole brief.
One more switch, and it is deliberately not in that table — it is not an .env key and it
changes nothing inside the container. Cloud Run scales this demo to zero when nobody is talking
to it, which is why an idle demo costs nothing, and why the first message after an idle gap
waits ~20-25s for a cold start and can come back as an error instead: Cloud Run sometimes
refuses a request outright while the container is still starting. Deploying with
MIN_INSTANCES=1 — the environment variable setup_and_deploy.sh reads in Step 4/6, exported
before the deploy runs — keeps one instance always up and removes both.
State it as a line under the table when you offer it, with the price attached rather than in a
footnote:
🔥 Warm instance — MIN_INSTANCES=1 · default 0 (scale-to-zero). No cold start and no
cold-start error on the first message. Cost: one 8 GiB / 2 vCPU instance is then billed
continuously for as long as the demo exists — not just during the presentation.
The honest default is 0: for a demo deployed now and shown next week, sending the first
message twice is far cheaper than a week of idle billing. It is also not a decision that has to
be made now — it is read at deploy time, so re-running the deploy with the variable set (or
unset) flips an existing demo either way.
§ The gate — ask once, then stop
Close the message with a single explicit question, in the demo's language, that (a) names what
approving will run — synthetic data generation, Google Drive upload, Cloud Run deploy, Gemini
Enterprise registration (Phases 3-7) — and (b) leaves the door open for late changes:
Shall I proceed with this design and run the synthetic data generation, Google Drive upload,
Cloud Run deployment and Gemini Enterprise registration (Phases 3-7) in one pass? Let me know
if you want any option enabled (Google Workspace OAuth, RAG mode, data scale, a warm Cloud
Run instance for a live session, …) or any part of the model changed.
One question, not a second round of Step 2.1: everything else was settled before the brief was
written, and a gate that re-opens five decisions is a gate nobody clears.
Then stop and wait. Do not start Phase 3 in the same turn, and do not treat "looks good" on
a previous message — the scenario choice in Phase 1 or an answer in Step 2.1 — as approval of
this brief.
When the user changes something, re-render the affected sections and ask again; an approved
brief is the specification the rest of the run is built from, so it has to be the version the
user actually said yes to.
Phase 3: Synthetic Data & External Sample Files Generation + Google Drive Upload
Entry condition: the Phase 2 brief was presented in full and the user approved it. If you
arrive here without that, go back and present it.
-
Derive the demo's identifiers, concrete name, and description:
SUFFIX=$(date +%s | tail -c 5)
DEMO_ID="${DOMAIN_SLUG}-${SUFFIX}"
SERVICE_NAME="ge-demo-${DOMAIN_SLUG}-${SUFFIX}"
DATASET_ID="demo_${DOMAIN_SLUG}_${SUFFIX}"
FIRESTORE_COLLECTION="demo-${DOMAIN_SLUG}-${SUFFIX}-tasks"
# Set concrete, domain-specific display name & description (matching Web UI / GAS version)
DEMO_DISPLAY_NAME="${COMPANY_NAME} ${AGENT_ROLE:-Operations Specialist}"
DEMO_DESCRIPTION="Orchestrates ${SCENARIO_SUMMARY:-operations and intelligent data analytics} across ${COMPANY_NAME}."
Always write DEMO_DISPLAY_NAME and DEMO_DESCRIPTION into .env so setup_and_deploy.sh and register_agent.py register the agent with concrete domain details.
-
Generate Real-World Synthetic Data & Display Previews:
- Generate CSV files with realistic addresses, valid foreign keys across tables, and varied statuses in
data/<table_name>.csv.
- Ensure all string data is generated in the detected target language.
- Run validation script:
python3 scripts/validate_csv.py data/*.csv.
- Render Dataset Preview & Record Counts: Output a Dataset Summary Table with total generated record counts for each BigQuery table, Firestore collection, and external Drive file, followed by 3–5 representative sample rows per table in clean Markdown preview tables (or
<carousel> sliders) clearly indicating the total row count in each table title.
- Write one line of grain per table into
data/<table_name>_description.txt ("one row per
POS transaction line", "one row per store"). The deploy folds it, the column
descriptions from data/<table_name>_schema.json, the row counts and the real date
range of every date column into adk_agent/app/data_assets.md — the DATA ASSET CATALOG
the agent's prompt is written around. Skipping the grain line costs only that line;
skipping the column descriptions costs the agent its schema.
2b. Amplify to Demo Volume (only when dataScale was agreed in Phase 2):
-
Generate External Sample Files & Upload Directly to Google Drive:
- Synthesize external demo files in
./external_files/:
- PDF Audit Report (
<domain>_audit_report.pdf): Multi-section structured document in target language with summary, details, and intentional 5-15% variance from BigQuery to trigger cross-source reasoning.
- Excel Spreadsheet Ledger (
<domain>_external_ledger.xlsx): Semi-structured workbook with localized KPI headers, units, and 40-80 transaction rows with FK references matching BigQuery tables.
- Simulated Operational Document Images (
handwritten_order_1.jpg, handwritten_order_2.jpg): Realistic scanned forms generated via gemini-3.1-flash-image with localized text.
- Upload to Google Drive:
uv run --no-project \
--with "openpyxl>=3.1.0,<4.0.0" \
--with "reportlab>=4.0.0,<6.0.0" \
--with "pillow>=10.0.0,<13.0.0" \
python3 scripts/generate_and_upload_external_files.py \
--domain "$DOMAIN_SLUG" \
--company "$COMPANY_NAME" \
--suffix "$SUFFIX" \
--outdir "./external_files" \
--spec-file "./data/external_files_spec.json"
--spec-file carries THIS demo's content (titles, sections, table rows, and the
style wording for the scanned forms), written in the demo's own language and
domain. Without it the script emits generic placeholder documents - it holds no
built-in industry content by design.
- Creates dedicated Drive folder:
GE Demo - <Company Name> (<Suffix>), reusing one
of that name if a previous run already made it.
- Uploads
.pdf, .xlsx, and .jpg files.
- Saves links to
external_files/drive_upload_summary.json and creates .url.json artifacts.
- The folder is owned by
${GCP_ACCOUNT} — the upload is a Drive v3 REST call
carrying this machine's own gcloud access token, so the account deploying the
demo owns what it creates and nothing has to be shared with it. Report that owner
and the folder URL. The script also asks for "anyone with the link (Reader)" as a
convenience; plenty of organizations refuse it, in which case the summary carries
share_error and the banner notes ℹ️ LINK SHARING OFF — the owner can still open
everything, so this is a footnote, not a failure.
Phase 4: ADK Multi-Agent Project Scaffolding & System Instructions
Scaffold the project in ./ge-demo-<domain>-<suffix>/.
Do not retype these files. Copy them from this skill's templates/ directory, which is
the single source of truth for every scaffolded file, then edit only the placeholders each
file marks. templates/ mirrors the tree below one-for-one (templates/agent.py ->
adk_agent/app/agent.py, templates/scripts/ -> scripts/, and so on):
ge-demo-<domain>-<suffix>/
├── adk_agent/
│ ├── __init__.py # empty; makes adk_agent importable as a package
│ └── app/
│ ├── __init__.py
│ ├── agent.py # Triple-Agent (gemini-3.8-flash) + Code Executor + A2UI System Prompts
│ ├── tools.py # MCP Toolsets + gemini-3.1-flash-image generate_image tool
│ ├── part_converters.py # A2A <-> Gen AI DataPart Converters
│ ├── fast_api_app.py # A2A Server + A2UI StreamParser + Token Middleware + /execute_task
│ ├── data_assets.md # DATA ASSET CATALOG (written at deploy time; do not hand-edit)
│ ├── catalogs/ # A2UI v0.9 Gemini Enterprise composite catalog
│ └── examples/0.9/ # A2UI few-shot example surfaces
├── viewer_app/ # Real-Time Operations Viewer Flask App
│ ├── main.py
│ └── requirements.txt
├── data/ # Generated CSV files + Knowledge Catalog schemas
├── external_files/ # Generated PDF, Excel & Scanned Images
├── scripts/
│ ├── managed_agent_instruction.txt # Autonomous sandbox agent's system instruction
│ ├── validate_csv.py # CSV formatting & schema auto-repair
│ ├── amplify_data.py # Grows hero CSVs to demo volume (deterministic, FK-safe)
│ ├── build_data_catalog.py # CSVs -> adk_agent/app/data_assets.md (the agent's schema)
│ ├── generate_and_upload_external_files.py # External files & Google Drive uploader
│ ├── setup_fs.py # Firestore seed documents
│ ├── setup_datastores.py # Discovery Engine DataStore creation & import
│ ├── register_agent.py # Gemini Enterprise agent registration
│ ├── create_managed_agent.py / warmup_managed_agent.py # Agent Engine managed agent
│ ├── dep_smoke_test.py # Pinned-dependency import smoke test
│ ├── preflight_check.py # Pre-deploy static checks
│ ├── verify_and_heal.py # Post-deploy verification & auto-repair
│ └── cleanup.sh # One-click teardown (Cloud Run, Agent Engine Sandbox, Managed Agent, BQ, FS, GCS)
├── .env # Configuration shared by setup_and_deploy.sh and cleanup.sh
│ # (full key reference: references/deployment_and_iam.md §3)
├── Dockerfile # Multi-stage container build
├── requirements.txt # Python dependencies
└── setup_and_deploy.sh # Standalone reproducible deployment script (strictly ordered)
Two files carry this demo's domain knowledge and BOTH must be filled in:
adk_agent/app/agent.py - the gen_instruction block, marked
TEMPLATE PLACEHOLDER. This is the conversational agent's business context.
scripts/managed_agent_instruction.txt - the same context for the autonomous
sandbox agent, in its [BUSINESS_CONTEXT] slot. Leave the file untouched and
enableManagedAgent - the one default-on capability - is silently skipped
at deploy time, taking delegate_autonomous_task and demo prompts 5 and 7
with it. [DATASET_ID] and [COLLECTION_ID] are substituted by
setup_and_deploy.sh; leave those alone. When a Workspace, computer-use or
operating-model option is on, the matching block has to be added too - see
references/multi_agent_architecture.md §4.
Critical A2UI System Instruction Requirements (v0.9 Composite Catalog)
- Output Placement Rule (Rule #0): Any text in the same response turn as a tool call is hidden. Only output
🔍 Analyzing... during tool execution. Final analytical reports, A2UI cards, and chips MUST appear in a separate turn with ZERO tool calls.
- Mandatory A2UI Cards (v0.9 Standard): Plain text markdown tables and bullet lists are forbidden for data presentation. Wrap results in
<a2ui-json> ... </a2ui-json> using MaterialCard, MaterialTable, VegaChart, MaterialRow, MaterialColumn, and MaterialDivider.
- A2UI v0.9 Component Model (CRITICAL):
- Every message stamped with
"version": "v0.9".
createSurface has NO root key: {"version": "v0.9", "createSurface": {"surfaceId": "...", "catalogId": "https://www.gstatic.com/vertexaisearch/a2ui/v0_9/gemini_enterprise_composite_catalog.json"}}.
- Root component MUST have
id: "root".
- Components use flat type names (
"component": "MaterialButton"), plain strings ("title": "..."), and plain arrays ("children": ["id1", "id2"]).
MaterialButton has a flat label: "..." and event action: {"action": {"event": {"name": "action_name", "context": {"prompt": "User message prompt"}}}}.
- Welcome Card on Greeting: On initial turn / greeting, DO NOT call tools. Output ONE line of plain-text greeting in user's language, followed by
<a2ui-json> onboarding card with surfaceId welcome-card and 3 action buttons. Do not emit suggestion chips on the welcome turn.
- Suggestion Chips: Append 3-4 follow-up
MaterialButtons with action.event.context.prompt at the end of every normal response, ALWAYS in their own trailing surface with surfaceId suggestions, emitted after the card — never as a MaterialRow inside the card. A turn's second A2UI surface does render; the rule that once said otherwise was wrong. Keeping the follow-ups out of the card keeps the answer card a clean read and makes the next actions a footer under it. (This layout is not a scroll fix — a press scrolls to the element of the user's PREVIOUS press, which no arrangement of surfaces can change; the server retires the pressed surface instead, see .) For the same readability reason a card carries no footer action row of its own — the exceptions are a button bound to its own card's fields (a binding only resolves within its surface) and the welcome card, whose buttons are its own content and which opens the conversation.
Phase 5: High-Speed & Resilient Deployment Sequence (Zero-Timeout)
Follow the optimized dependency sequence with local pre-flight checks and fast builds:
-
Step 0: Local Pre-flight Verification (1-Second Syntax & Import Check):
Prevents 5-minute remote Cloud Build / Cloud Run startup health check timeouts by catching syntax/import errors instantly.
python3 scripts/preflight_check.py
-
Step 2/6: Provision Agent Engine Sandbox (Code Execution Environment):
Must be executed from a clean temporary directory (mktemp -d) to prevent SDK build hangs.
The heredoc runs in a child process, so the variables it reads must be exported — a plain
source .env leaves them shell-local and every os.environ.get() below silently returns the default.
set -a; [ -f .env ] && source .env; set +a
export PROJECT_ID SERVICE_NAME
SANDBOX_TMPDIR=$(mktemp -d)
pushd "$SANDBOX_TMPDIR" > /dev/null
python3 - << '__SANDBOX_EOF__'
import os, vertexai
from vertexai import types
client = vertexai.Client(project=os.environ.get('PROJECT_ID', ''), location='us-central1')
ae = client.agent_engines.create(config={'display_name': os.environ.get('SERVICE_NAME', 'demo') + '-sandbox'})
sb = client.agent_engines.sandboxes.create(
name=ae.api_resource.name,
config=types.CreateAgentEngineSandboxConfig(display_name='code-sandbox'),
spec={'code_execution_environment': {}}
)
with open('/tmp/sb_out.txt', 'w') as f:
f.write(f"{ae.api_resource.name}|{sb.response.name}")
__SANDBOX_EOF__
popd > /dev/null
rm -rf "$SANDBOX_TMPDIR"
AGENT_ENGINE_NAME=$(cat /tmp/sb_out.txt | cut -d'|' -f1)
SANDBOX_RESOURCE_NAME=$(cat /tmp/sb_out.txt | cut -d'|' -f2)
-
Step 3/6: Deploy Data Viewer Dashboard (Cloud Run with --no-allow-unauthenticated + IAP):
The service name MUST be ge-viewer-${SERVICE_NAME}. scripts/cleanup.sh reconstructs it
from that exact expression; any other name leaves the viewer running and billing after teardown.
VIEWER_SERVICE_NAME="ge-viewer-${SERVICE_NAME}"
gcloud run deploy "$VIEWER_SERVICE_NAME" \
--source viewer_app \
--region "$REGION" \
--platform managed \
--ingress all \
--no-allow-unauthenticated \
--set-env-vars="PROJECT_ID=${PROJECT_ID},FIRESTORE_COLLECTION=${FIRESTORE_COLLECTION},DEMO_ID=${DEMO_ID},DASHBOARD_TITLE=${DOMAIN_SLUG} Operations Console,SYSTEM_DESCRIPTION=Real-Time Operational Intelligence Dashboard"
# Enable IAP and grant deployer access
PROJECT_NUMBER=$(gcloud projects describe "$PROJECT_ID" --format="value(projectNumber)")
gcloud beta services identity create --service=iap.googleapis.com --project="$PROJECT_ID" >/dev/null 2>&1 || true
gcloud run services add-iam-policy-binding "$VIEWER_SERVICE_NAME" --region="$REGION" --member="serviceAccount:service-${PROJECT_NUMBER}@gcp-sa-iap.iam.gserviceaccount.com" --role="roles/run.invoker" --project="$PROJECT_ID" >/dev/null 2>&1 || true
gcloud beta run services update "$VIEWER_SERVICE_NAME" --region="$REGION" --iap --project="$PROJECT_ID"
DEPLOYER_EMAIL=$(gcloud config get-value account 2>/dev/null)
if [ -n "$DEPLOYER_EMAIL" ]; then
gcloud beta iap web add-iam-policy-binding --project="$PROJECT_ID" --resource-type=cloud-run --region="$REGION" --service="$VIEWER_SERVICE_NAME" --member="user:$DEPLOYER_EMAIL" --role="roles/iap.httpsResourceAccessor" >/dev/null 2>&1 || true
fi
VIEWER_URL=$(gcloud run services describe "$VIEWER_SERVICE_NAME" --region="$REGION" --format="value(status.url)")
Phase 6: Gemini Enterprise App Discovery, Authorization & Registration (Step 6/6)
-
Check Discovery Engine / Gemini Enterprise App Existence:
- Query Discovery Engine API across regions (
global, us, eu).
- If 1 GE App found: select it automatically.
- If 0 GE Apps found: automatically create
default-gemini-enterprise-app via Discovery Engine API.
-
Google Workspace OAuth Authorization Linking (when either enableWorkspaceAuth or enableWorkspaceMcp is on):
- Check Secret Manager for
ge-demo-oauth-client-id and ge-demo-oauth-client-secret.
- If missing, guide user through step-by-step setup in Console and save credentials.
- Create Discovery Engine Authorization resource with stored credentials.
- Pass
--authorization-id=projects/$PROJECT_ID/locations/global/authorizations/$AUTH_ID during registration.
-
Register Agent to Gemini Enterprise:
# Discovery Engine Service Account permissions
gcloud run services add-iam-policy-binding "$SERVICE_NAME" \
--region="$REGION" \
--member="serviceAccount:service-${PROJECT_NUMBER}@gcp-sa-discoveryengine.iam.gserviceaccount.com" \
--role="roles/run.servicesInvoker"
# Register via agents-cli
agents-cli publish gemini-enterprise \
--agent-card-url "${SERVICE_URL}/a2a/app/.well-known/agent-card.json" \
--display-name "${COMPANY_NAME} Demo Agent" \
--description "${DEMO_DESCRIPTION}" \
${AUTH_FLAG}
Phase 6.5: Autonomous Verification & Real-Time Self-Healing Engine (Mandatory Gate)
Immediately after deployment and registration complete, execute the automated 8-layer verification and self-healing engine before presenting results to the user:
python3 scripts/verify_and_heal.py
The 8-Layer Autonomous Verification & Healing Matrix
- Layer 1: BigQuery Schema & Primary Keys: Inspects all tables in
${DATASET_ID}, verifies row counts, and auto-heals missing _id document columns (ALTER TABLE ... ADD COLUMN IF NOT EXISTS _id STRING; UPDATE ... SET _id = <PK>) to guarantee Discovery Engine DataStore ingestion compatibility.
- Layer 2: Firestore Operations Collection: Verifies operational task documents count >= 3. Auto-seeds initial documents if missing.
- Layer 3: Data Viewer & IAP: Verifies Cloud Run dashboard service status and guarantees
roles/run.invoker binding for service-${PROJECT_NUMBER}@gcp-sa-iap.iam.gserviceaccount.com.
- Layer 4: Agent Engine Sandbox: Confirms Agent Engine Code Sandbox resource accessibility.
- Layer 5: Cloud Run A2A & Fallback Routing: Probes
/openapi.json, /a2a/app/.well-known/agent-card.json, and root POST fallback alias @app_instance.post("/").
- Layer 6: Discovery Engine DataStores & Engine Binding: Inspects indexed document count in
ds-<service>-bq and ds-<service>-gcs, auto-restarts table ingestion if 0 documents, and verifies dataStoreIds attachment on the Gemini Enterprise Assistant Engine.
- Layer 7: Agent Registry URL & Authorization: Verifies registered agent URL strictly ends with
/a2a/app, auto-patches Gemini Enterprise agent card if missing, and verifies Authorization resource formatting (projects/${PROJECT_NUMBER}/...). Resolves direct chat link https://vertexaisearch.cloud.google.com/home/cid/${CONFIG_ID}/r/agent/${AGENT_ID}/session/-.
- Layer 8: External Files & Google Drive: Verifies external PDF, Excel, and Scanned Image files staging.
Phase 7: Comprehensive Results Output & 7 Demo Prompts Playbook
Upon deployment completion, ALWAYS output the structured results containing direct links and the 7 Structured Demo Prompts Playbook localized into the customer's target language:
1. Deployment Identity & Direct Links Section
### 👤 Deployment Identity & Environment
- 👤 **Deployed By Account**: `${GCP_ACCOUNT}`
- 🏢 **Target Project**: `${PROJECT_ID}` (Project Number: `${PROJECT_NUMBER}`)
- 🌐 **Deployed Region**: `${REGION}`
- 🤖 **Cloud Run Service Account**: `${PROJECT_NUMBER}-compute@developer.gserviceaccount.com`
---
### 🔗 Quick Access Links
> [!IMPORTANT]
> **Account Notice:**
> Open these links in a browser whose Google Cloud / Workspace session is the deploying account **`${GCP_ACCOUNT}`**. Opening them as a different account returns a permission error (403 Forbidden).
> Write this hand-off summary in the language the user is conversing in - the wording above is the English form, not a fixed string.
💬 **Start Chatting with Your Agent (Direct Chat Link):**
👉 https://vertexaisearch.cloud.google.com/home/cid/${CONFIG_ID}/r/agent/${AGENT_ID}/session/-
⚠️ The FIRST message right after a deploy can come back as an error. Gemini Enterprise takes up to ~90 seconds to start routing to a brand new agent, and a cold Cloud Run instance can refuse one request while it starts. Both are transient — send the message again.
💻 **Gemini Enterprise Console (Overview):**
👉 https://console.cloud.google.com/gemini-enterprise/locations/${SELECTED_LOC}/engines/${SELECTED_APP_ID}/overview/dashboard?&project=${PROJECT_ID}
*(Or fallback console if no app registered: `https://console.cloud.google.com/gemini-enterprise/overview?&project=${PROJECT_ID}`)*
📁 **Google Drive External Sample Files** (when the deploy-time upload ran):
- 👑 **Folder Owner**: `${DRIVE_OWNER_ACCOUNT}` (the account that ran the deploy — switch
the browser to it before opening these links)
- 🔑 **Link Sharing**: "Anyone with link (Reader)" when the organization allows it
- 📂 **Open the Folder**: https://drive.google.com/drive/folders/${DRIVE_FOLDER_ID}
- 📄 Audit Report (PDF, Japanese CJK Font): ${PDF_URL}
- 📊 Supplier Ledger (Excel): ${XLSX_URL}
- 🖼️ Simulated Document 1 (JPG): ${IMG1_URL}
- 🖼️ Simulated Document 2 (JPG, Discrepancy Embedded): ${IMG2_URL}
- ℹ️ If `drive_upload_summary.json` carries `share_error`, link sharing was refused: say
the owner can open the folder but the audience needs an explicit share from the Drive
UI, and offer the Cloud Storage links below in the meantime.
📦 **External Sample Files in Cloud Storage** (always — the staging copy is unconditional):
- 🗂️ **Browse the bucket**: https://console.cloud.google.com/storage/browser/${GCS_BUCKET_NAME}?project=${PROJECT_ID}
- 📄 **Open a file directly** (signed-in browser): one line per uploaded file,
`https://storage.cloud.google.com/${GCS_BUCKET_NAME}/<filename>` — print the real
filenames, never a bare `gs://` URI on its own; `gs://` is not clickable.
- 📂 Also on this machine: `./external_files/`
📁 **When the Drive upload was skipped** (the credentials carry no Drive scope) — say it
plainly, this is the only notice the user gets:
- ℹ️ **There is no Drive copy of these documents.** Print `upload_skipped_reason` verbatim.
The documents themselves are complete, in Cloud Storage and `./external_files/`; only the
Drive folder is missing, so a Drive or Sheets step in the demo script will find nothing.
- 🛠️ **To get one**: run `gcloud auth login --enable-gdrive-access --no-launch-browser` and re-run this script
(a plain `gcloud auth login` grants no Drive scope, which is why this is the usual
cause; `--no-launch-browser` prints the OAuth consent URL to copy into any browser when the terminal environment lacks a browser), or upload `./external_files/` to a Drive folder by hand and share it with the
audience. Never suggest asking the agent to do it — it cannot.
📊 **Firestore Data Viewer Dashboard:** 👉 ${VIEWER_URL}
🔎 **BigQuery Console:** 👉 https://console.cloud.google.com/bigquery?referrer=search&project=${PROJECT_ID}&ws=!1m4!1m3!3m2!1s${PROJECT_ID}!2s${DATASET_ID}
2. The 7 Structured Demo Prompts Playbook (Localized)
Present all 7 demo prompts formatted with Title/Persona, Category Tags, Copyable Prompt Text,
Expected Outcome, and a one-sentence Watch Point for the person running the demo (what to
look at on screen while this prompt runs). All 7 trace the one process instance from Phase 2,
and the last one closes with quantified outcomes — before/after cycle time, items resolved,
hand-offs completed.
Slot 5 and slot 7 carry the two autonomous delegation prompts whenever enableManagedAgent
is on (the default), which moves the interactive-dashboard prompt into slot 1 or 2. If the
enabled capabilities still do not fit, append up to three prompts tagged encore rather than
cramming two showcases into one prompt. Both rules, and the per-capability chains for
Workspace and computer use, are in references/demo_prompts_guide.md §5-§6.
The template below is the base progression, before those overrides:
### 🎯 7 Structured Demo Prompts Playbook (Dynamically Localized to Target Language)
#### 1. [Role Title] Foundation & Data Overview
- **Tags**: `[Foundation]` `[Data Overview]`
- **Prompt Text**: (Generically phrased request to explore data landscape and operational KPIs in target language)
- **Expected Outcome**: Analyzes master/transaction tables and renders KPI summary in an A2UI Card.
- **Watch Point**: (What the operator should notice - e.g. the console already shows items mid-process across departments)
#### 2. [Role Title] Metadata & Knowledge Catalog Discovery
- **Tags**: `[Metadata Discovery]` `[Knowledge Catalog]`
- **Prompt Text**: (Generically phrased request inquiring about available data resources, metrics definitions, and relationships)
- **Expected Outcome**: Consults Knowledge Catalog MCP (`search_entries`, `lookup_entry`) before writing queries.
- **Watch Point**: (e.g. the agent reads the catalog before it writes a single query)
#### 3. [Role Title] Cross-Source Anomaly & Risk Detection [WOW MOMENT]
- **Tags**: `[Cross-Source WOW]` `[Drive File Binding]`
- **Prompt Text**: (Strategic inquiry about untracked discrepancies across recent deliveries/records)
- **Expected Outcome**: Autonomously cross-references the external PDF/Excel against BigQuery tables, isolates the 5-15% discrepancy, and renders discrepancy cards and infographics.
- **Watch Point**: (e.g. nobody told it to open the external report - it decided to)
#### 4. [Role Title] Multi-Step Dependent Immediate Workflow [WOW MOMENT]
- **Tags**: `[Immediate Workflow WOW]` `[A2UI Batch Editor]`
- **Prompt Text**: (Request to scan unverified items, resolve mappings, and update records)
- **Expected Outcome**: Executes `SCAN -> RESOLVE -> PRESENT -> EXECUTE -> AUDIT`, presenting the (J) Dynamic Multi-Entity Batch Editor A2UI form for human confirmation before writing to DB.
- **Watch Point**: (e.g. after the approval click, the item moves to the next department on the operations console)
#### 5. [Role Title] Large-Scope Batch / Background Reconciliation [WOW MOMENT]
- **Tags**: `[Background Workflow WOW]` `[Execution Mode Dialog]`
- **Prompt Text**: (Comprehensive quarterly reconciliation request across all historical records)
- **Expected Outcome**: Recognizes large batch scope and presents Execution Mode Dialog (Immediate vs Background vs Scheduled), kicking `/execute_task` when background mode is selected.
- **Watch Point**: (e.g. the agent proposes background mode on its own, then keeps the chat usable while it runs)
#### 6. [Role Title] Scheduled Automated Monitoring Setup
- **Tags**: `[Scheduled Monitoring]` `[Pub/Sub Task]`
- **Prompt Text**: (Request to set up automated recurring threshold monitoring every morning at 09:00 AM)
- **Expected Outcome**: Explains monitoring logic, registers recurring cron schedule with Pub/Sub.
- **Watch Point**: (e.g. the schedule it proposes matches the department's own escalation rule)
#### 7. [Role Title] End-to-End Strategic Automation
- **Tags**: `[Strategic Automation]` `[End-to-End]`
- **Prompt Text**: (Comprehensive executive request combining cross-source analytics, workflow execution, notification drafting, and audit logging)
- **Expected Outcome**: Synthesizes all data sources, produces executive summary infographic, updates records, and logs audit trail.
- **Watch Point**: (e.g. the closing summary states the before/after cycle time for the instance the whole demo followed)
Phase 8: Cleanup & Teardown (On-Demand)
When the demo is concluded, delete all provisioned resources (including the Agent Engine
Sandbox and the Managed Autonomous Agent):
bash scripts/cleanup.sh
Read the per-resource lines, not the closing banner. Every job runs under || true so that one
failure cannot strand the rest, which means the script finishes whatever happened: ✅ deleted,
⚠️ already gone or skipped, ❌ still there. A ❌, or a ⚠️ for something you know existed,
needs a manual delete — an Agent Engine, a bucket or a Firestore collection left behind keeps
billing. Run it from the demo directory so it picks up .env; without DOMAIN_SLUG/SUFFIX it
cannot name the two GCS buckets and says so rather than guessing.
References & Templates
- For Discovery Engine DataStore Connectors & Search: read
references/datastore_connectors.md.
- For Deployment sequence, Dependency graph & the full
.env key reference: read references/deployment_and_iam.md.
- For External sample files & Google Drive upload: read
references/external_files_and_drive.md.
- For 7 Demo prompts specification: read
references/demo_prompts_guide.md.
- For Multi-Agent routing & Sandbox code execution: read
references/multi_agent_architecture.md.
- For MCP server integrations & Workspace Auth: read
references/mcp_catalog.md.
- For Data generation & Knowledge Catalog metadata: read
references/data_generation.md.
- For A2UI JSON payload format & system instructions: read
references/a2ui_catalog.md.