Vertex AI as the AI platform

The fourth session moved down the stack. Jules arrived with a printed sketch of the shopping assistant: a chat window in the corner of birchline.example, a customer typing "I'm doing a three-day canoe trip in Algonquin in September, what do I need?" and the assistant answering with a packing list, links to products, and a note that the 40-litre dry bag was in stock at the nearest store. "None of the products we've looked at so far can do this," he said. "It has to be ours."

"Then it gets built on the platform," Farah said. "The lesson title in your course still says Vertex AI, because that's the name most people know. Google relaunched it in April 2026 as the Gemini Enterprise Agent Platform. Same purpose, more agent tools, new name. When you read Vertex AI in older material, read Agent Platform." This lesson explains what the AI platform is, what it contains, who uses it, and why a company builds there rather than stitching together separate tools.

From Vertex AI to Gemini Enterprise Agent Platform

Vertex AI launched in 2021 as Google Cloud's unified machine learning platform: one place to prepare data, train models, deploy them and monitor them. As generative AI grew, it added access to Gemini and other foundation models, prompt design tools, grounding, tuning and agent building. In April 2026, at Google Cloud Next, Google relaunched the platform as the Gemini Enterprise Agent Platform, describing it as the successor to Vertex AI with a stronger focus on building, deploying and governing agents. Many feature names changed with it: for example, the exam guide now refers to Agent Search, Agent Platform AutoML, Agent Platform Feature Store and Agent Studio.

For a leader the change matters less than it seems. The platform's job is the same: give builders a managed, secure place to use models and turn them into applications and agents. The exam uses the new names, so learn them, but expect to meet the old names in documentation, job postings and colleagues' conversations for some time.

What the platform contains

Farah grouped the platform's capabilities by the job a builder is doing, because that is how people use it and how exam scenarios describe it.

Who uses the platform

The platform serves several roles, each at a different depth. Developers call models through APIs and build applications and agents. Data scientists and machine learning engineers train, tune and evaluate models and manage their lifecycle. Business analysts may use low-code tools such as Agent Studio or AutoML. Platform and security teams set up projects, access, networking and policies so that everyone else builds safely. Leaders rarely use the platform themselves, but they decide whether a project belongs there and what skills it needs.

At Birchline the shopping assistant team would be small: two developers from the web team, Ravi's data engineer for the product catalogue, Teresa for security review, and Jules as product owner. Farah would design the architecture. Owen's role was to understand it well enough to explain choices to the steering committee.

Predictive AI and generative AI on one platform

The platform is not only for generative AI. It also supports traditional predictive machine learning, in which a model learns from labelled examples to predict a number or category, such as next month's demand for sleeping bags or whether an order is likely to be returned. GCP-110 explained the difference. Having both on one platform matters because real solutions often combine them: Birchline's demand forecast, already built with BigQuery ML in GCP-102, could feed the shopping assistant so it avoids recommending items about to run out.

Agent Platform AutoML lets a team train a predictive model on its own tabular, image or text data with little code: the team supplies labelled data and AutoML tries model types and settings, then reports how well the best model performs. It suits teams with good data but no machine learning specialists. Custom training, where data scientists write their own training code, suits unusual problems or teams that need full control.

Adapting a foundation model

Most generative AI projects start with a foundation model as it comes, plus careful prompting and grounding. When that is not enough, the platform offers ways to adapt a model, in rising order of effort.

Farah's advice echoed GCP-110: try prompting first, add grounding when facts matter, and only consider tuning when there is good evidence that the first two cannot meet the need. Tuning needs a set of high-quality examples, costs money to run, and produces a model version that must be evaluated and maintained. For the shopping assistant, grounding in the product catalogue mattered far more than tuning, because products and stock change every day.

Evaluating before and after launch

How does a team know the assistant is good enough to put in front of customers? The platform includes evaluation tools that score a model's or an agent's answers against criteria such as accuracy, relevance, groundedness in the sources provided, safety and following instructions. Some scores compare an answer with a reference answer; others use a second model to judge qualities that are hard to measure automatically. Evaluation runs before launch, to choose between prompts or models, and continues afterwards, to catch quality slipping when products, policies or the model change.

# Illustrative evaluation summary for two prompt versions of the shopping assistant.
# Each score is the share of 200 test questions that reviewers marked acceptable (out of 1.0).
results = {
    "Prompt v1": {"accurate": 0.86, "grounded": 0.81, "on brand": 0.92, "safe": 0.995},
    "Prompt v2": {"accurate": 0.91, "grounded": 0.93, "on brand": 0.90, "safe": 0.995},
}
weights = {"accurate": 0.4, "grounded": 0.3, "on brand": 0.1, "safe": 0.2}
for name, scores in results.items():
    total = sum(scores[k] * w for k, w in weights.items())
    print(f"{name}: weighted score {total:.3f}")
    for k, v in scores.items():
        print(f"   {k:9} {v:.3f}")

Prompt v1: weighted score 0.878
   accurate  0.860
   grounded  0.810
   on brand  0.920
   safe      0.995
Prompt v2: weighted score 0.932
   accurate  0.910
   grounded  0.930
   on brand  0.900
   safe      0.995

The weights in this illustrative example reflect Jules's priorities: accuracy and groundedness matter most, and safety has to be near perfect in every version regardless of its weight. Version 2 gives more grounded answers at a small cost in brand tone, which the team could fix with an extra instruction. The point for a leader is that model and prompt choices should be made with numbers from a fixed test set, not from a handful of impressive demos.

Deploying and operating

Once a model or agent is ready, the platform runs it as a managed service. Google's foundation models are called through APIs and scale automatically; Birchline pays for what it uses, typically by the amount of input and output processed. Custom or open models can be deployed to managed endpoints, where the team chooses machine types and scaling. Agents run on a managed agent runtime. In every case the platform provides monitoring, logging and quotas, which plug into the Cloud Monitoring and Cloud Logging tools GCP-101 introduced.

gcloud services enable aiplatform.googleapis.com --project=bl-shop-dev
gcloud ai models list --region=REGION --project=bl-shop-dev

The API name in the command above still contains aiplatform, a reminder that product names on the surface change faster than the plumbing underneath.

Why build on the platform rather than piece it together

Jules asked why Birchline could not simply call a model API from the web server and skip the platform. Farah gave four reasons that also appear in exam scenarios.

• Security and governance come with it. The platform runs inside Birchline's projects, under its IAM roles, organization policies, audit logs and network controls.
• The pieces work together. Model access, grounding, evaluation, agent runtime and monitoring share data and settings, so the team builds less glue code.
• Choice without lock-in to one model. Model Garden lets the team switch or compare models as better or cheaper ones appear.
• Enterprise data commitments. Google states that customer data used with the platform's models is not used to train Google's models without permission, and the platform supports controls such as data residency and customer-managed encryption keys, covered in lesson 10.

Projects, environments and the platform

Platform work follows the same structure GCP-101 set up. The shopping assistant would be built in bl-shop-dev, tested there against copies of the catalogue, and promoted to bl-shop-prod only after evaluation and security review. Each project has its own quotas for model requests, its own budget alerts and its own audit logs, so experiments in development cannot exhaust production capacity or hide in production costs. Dana asked for a separate budget line for AI usage in each project, so the finance team could see how much the assistant cost per conversation as it grew.

By the end of the session Owen could draw the platform's six jobs from memory. Jules left with a list of the skills his team would need, and a note to ask the web developers whether they had used the Agent Development Kit.