Agents are where gen AI shifts from answering questions to getting work done. An agent can call tools, look up data, and take multi-step actions toward a goal. For a leader, agents unlock automation of real workflows -- but they also raise new questions about tool permissions, oversight, and reliability. Vertex AI Agent Builder is Google Cloud's way to build these systems without assembling everything by hand.

An AI agent uses a model as a reasoning engine to plan and act, not just to generate text. The core mechanism is tool use, often via function calling: the model decides it needs an action, outputs a structured call, your system runs the tool (an API, a search, a database query), and the result is fed back so the agent continues. Orchestration coordinates these steps, memory, and multiple tools toward the goal.

Vertex AI Agent Builder is a managed offering to create agents grounded on your data and connected to tools, with the platform handling much of the orchestration, grounding, and deployment. It lets teams stand up assistants and task automation on Google Cloud without building the agent loop from scratch.

An operations team wants an assistant that checks order status and can start a refund. A pure chatbot cannot act. As an agent, the model is given two tools: a status-lookup API and a refund API. When a customer asks about a late order, the agent calls the lookup tool, reads the result, and -- only within policy limits and with human approval for large refunds -- calls the refund tool. Agent Builder wires the tools, grounds the agent on policy docs, and enforces that high-value actions route to a human. This blend of autonomy plus guardrails is the exam's ideal.

Agents build on everything prior: the model is a foundation model on Vertex AI, tools often include RAG for grounded answers, and the risk of an agent taking a wrong action is exactly why Responsible AI, permission scoping, and human-in-the-loop (next topic) matter most here.