Cheat sheetGCP-04

AI Agents & Agent Builder

GCP GenAI Leader / AI Agents & Agent Builder

Agents act, not just answer: a model plans and calls tools (via function calling), coordinated by orchestration. Agent Builder makes this managed.

Agent vs chatbot
Action-takingAn agent uses tools to complete tasks; a chatbot only produces text.
Function calling
Tool mechanismModel emits a structured call; your system runs it and returns the result to continue.
Agent Builder
Managed agentsVertex AI Agent Builder grounds agents on your data and wires tools, orchestration, and deployment.
Guardrails
Scope + approvalLeast-privilege tool access and human approval on high-impact actions.
⚑

Give agents only the tools they need, and gate consequential actions (refunds, deletes, sends) behind human approval.

Good agentOrder-status lookup plus in-policy refund tool, with large refunds sent to a human.
Risky agentOne agent with unscoped access to every internal API and no approval gates.
Agent = model plans + calls tools; chatbot = text only.
Function calling: model -> structured call -> run tool -> result back.
Agent Builder = managed grounded agents; still keep humans on high-impact steps.
agentsfunction-callingagent-builderorchestrationhuman-oversight
review in 6d