Vertex AI (now Gemini Enterprise Agent Platform)

1 · What it is

Vertex AI is Google Cloud’s service for building with AI models. Google Cloud now calls it Gemini Enterprise Agent Platform and presents it as the next step for Vertex AI.

The platform does four jobs. First, it helps you pick a model. Google says it offers more than 200 foundation models (large general-purpose models that others build on). The catalogue, called Model Garden, lists Google’s own models (Gemini, Gemma, Veo, Lyria). Partner models on the list include Anthropic’s Claude, Grok and Mistral AI models, and there are open-weights options (models whose trained numbers anyone can download) such as Qwen, Llama and DeepSeek. Companies can also set a rule that allows only models they have checked.

Second, it helps you tune a model. Tuning means training an existing model a little more on your own examples, so it gets better at one narrow job. Google’s advice is to try prompting first. Fine-tuning works best with a sizeable labelled set, ideally 100 examples or more. A labelled example is an input paired with the answer you want, such as a review paired with “unhappy”. Parameter-efficient tuning changes only a fairly small part of the model’s parameters, the numbers that shape its answers. Full fine-tuning changes every parameter.

Third, it runs models for real traffic. There are two ways to reach a model. You can use a managed API (Google calls this Model as a Service), or deploy the model yourself. With a managed API, you send a request and get an answer back, and you never run a server. Deploying means linking a model to its own machines so it can answer quickly, and giving it an endpoint, a web address your app sends requests to. You are charged for the machines behind a deployed open model. In return you get controls. Autoscaling adds or removes serving machines based on how many requests arrive at once. A new model version can share an endpoint with the old one. It starts with a small share of traffic and gets more until it handles all of it.

Fourth, it runs agents. Agent Runtime can run agents built with Agent Development Kit, a ready-made toolkit for building agents. Agents on Agent Runtime can use Sessions and Memory Bank to keep track of ongoing conversations and long-term memories.

Before adding more training examples, Google advises checking where the model goes wrong. For agents, Google offers an Example Store and an Evaluation Service. They help you test, monitor and trace what an agent does, and improve it over time.

Picture a small online shop that wants every review tagged with its own labels. The team tries a Gemini model from Model Garden with a clear prompt. If the labels keep drifting, they collect a few hundred reviews that staff have already tagged and tune the model on them. They put the tuned version on an endpoint, and their website sends each new review there. Later, they give a better version a small slice of reviews first.