Introduction to Gemini Enterprise Agent Platform

Note: This lab pre-provisions a Windows 11 VM and leases a Google Cloud project at start — allow up to 12 minutes for the environment to become ready before beginning the exercises.

Make your first Gemini call from Python using the unified google-genai SDK pointed at the Gemini Enterprise Agent Platform. You will sign into Antigravity on the Windows VM, open the pre-staged hello-gemini starter in the Antigravity IDE, enable the Agent Platform API, configure Application Default Credentials for the lab's workforce identity, complete three TODOs in main.py, and compare the Gemini 3.1 Flash, Gemini 3.1 Pro, and Gemini 2.5 Flash-Lite models on a real Nimbus Outfitters merch-desk prompt. By the end of the lab you will be able to initialize a google-genai client against Gemini Enterprise Agent Platform, send a prompt, read the response's usage metadata and finish reason, and reason about which Gemini model fits a given task.

Note: This lab pre-provisions a Windows 11 VM and leases a Google Cloud project at start — allow up to 12 minutes for the environment to become ready before beginning the exercises.

Put Gemini Enterprise Agent Platform auth into practice on a Windows VM with the Antigravity IDE. You will sign into Antigravity, open the auth-lab starter, inspect a pre-provisioned merch-desk-sa service account with its single least-privilege roles/aiplatform.user binding, run the starter's verify.py under three different credential setups (your workforce user, a service-account key file, keyless impersonation via --impersonate-service-account), deliberately break the auth three ways to recognize DefaultCredentialsError, PermissionDenied, and FailedPrecondition on sight, and finish with a fully key-less posture. By the end of this lab you will be able to pick the right auth pattern for a Gemini workload, read any Gemini Enterprise Agent Platform auth traceback and name the fix in one sentence, and argue why service-account impersonation beats downloaded JSON keys for production.

Note: This lab pre-provisions a Windows 11 VM and leases a Google Cloud project at start — allow up to 12 minutes for the environment to become ready before beginning the exercises.

Build a Nimbus Outfitters review-summarizer CLI from a Python starter on the Windows VM in the Antigravity IDE. You will sign into Antigravity, open the pre-staged review-summarizer starter, wire up a single-call Gemini summarizer across 20 product reviews, switch the path to live streaming, track per-review token counts, benchmark Gemini 3.1 Flash against Gemini 3.1 Pro and Gemini 2.5 Flash-Lite to pick a production model, and extend the CLI with a multi-turn chat subcommand that holds conversation history across follow-up questions. By the end of the lab you will know when to pick streaming vs non-streaming on the Gemini Enterprise Agent Platform, how to read usage_metadata off both branches, how to defend a model choice to a non-engineering stakeholder with real numbers, and how client.chats.create differs from stateless generate_content calls.

Note: This lab pre-provisions a Windows 11 VM and leases a Google Cloud project at start — allow up to 12 minutes for the environment to become ready before beginning the exercises.

Build a Python CLI that feeds Nimbus Outfitters product photos through Gemini on the Gemini Enterprise Agent Platform and produces structured catalog entries. You will sign into Antigravity on the Windows VM, open the pre-staged photo-describer starter, create a regional Cloud Storage bucket in us-central1, upload five sample product photos, call Gemini with Part.from_uri for multimodal input, add JSON Schema–constrained structured output with response_schema, harden the pipeline against non-image inputs with a MIME allowlist, and close the loop by generating brand-new catalog hero images with Gemini 3.1 Flash Image (and compare against Gemini 3.1 Pro Image). By the end of the lab you will be able to compose multimodal contents with mixed image + text Parts, pin Gemini output to a JSON schema, defend a loop against bad inputs, and reason about when to reach for the Pro image-generation tier versus the GA Flash tier.

Note: This lab pre-provisions a Windows 11 VM and leases a Google Cloud project at start — allow up to 12 minutes for the environment to become ready before beginning the exercises.

Wire a two-document system instruction (tone_of_voice.md + content_policy.md) into Gemini via GenerateContentConfig.system_instruction, emit structured CatalogCopy JSON validated with Pydantic, deliberately flip a safety threshold to force a finish_reason=SAFETY block and handle it cleanly, enable thinking mode on gemini-3.1-pro-preview to measure thoughts_token_count on a complex SKU, and harden the system instruction against prompt-injection briefs. You will work in the Antigravity IDE on a Windows 11 VM with the pre-staged copy-drafter starter already on disk. By the end of this lab you will be able to compose multi-document system instructions, enforce structured output with a Pydantic schema, read safety ratings off a blocked response, make a cost-vs-quality judgment on thinking mode, and audit your output for policy leakage.