What's changed: Created Generative AI Leader Chapter 2 (Domain 2 "Google Cloud's generative AI offerings": Gemini and models/Vertex AI = Gemini/Imagen/Veo/Chirp/Model Garden/Vertex AI/Vertex AI Studio; gen AI embedded in work = Gemini for Workspace/Gemini for Google Cloud/NotebookLM/Conversational Agents (Customer Engagement Suite); agents and grounding = agentic AI/Vertex AI Agent Builder/Grounding/AI infrastructure (TPU/GPU)).
2.3Agents and grounding
Understand agents (agentic AI) that autonomously execute tasks toward a goal and Vertex AI Agent Builder to build them, grounding that bases answers on Google Search or your own data, and the role of the AI infrastructure (TPUs/GPUs) that powers generative AI—from a business perspective.
AI that does more than return text—autonomously planning steps, using tools, and taking actions toward a goal—is called an agent (agentic AI). For example, given "arrange my business trip," it performs a sequence of actions: search, book, and add to the calendar. Agents are a key direction for applying generative AI.
2.3.1Vertex AI Agent Builder and grounding
Vertex AI Agent Builder is the means to build such agents and search apps. The key to useful agents is grounding—basing the model answers on trusted sources. For recency, use Grounding with Google Search (based on Google Search); for internal information, ground on your own data (RAG, Chapter 3). Grounding reduces hallucinations and lets you cite sources, so it is especially important for business use.
2.3.2AI infrastructure
Training and inference for generative AI require enormous compute, and Google provides AI infrastructure such as purpose-built TPUs (Tensor Processing Units) and GPUs. Leaders need not know the details, but should grasp that "large-scale AI is powered by specialized high-performance infrastructure" and that this relates to cost and sustainability (power). Google provides this infrastructure in a managed way through Vertex AI and more.
Agents call external tools and services to do real work. As the foundation, they combine Cloud Storage and databases for storing data, Cloud Functions for lightweight event-driven processing, and Cloud Run for running containerized custom logic serverless. As a leader, it is enough to picture that "behind an agent, serverless runtimes like Cloud Run and Cloud Functions run only when needed to do the work."
| Concept | Meaning |
|---|---|
| Agent (agentic AI) | Autonomously plan, use tools, and act toward a goal |
| Vertex AI Agent Builder | Build agents and search apps |
| Grounding | Base answers on trusted sources |
| AI infrastructure (TPU/GPU) | Specialized hardware powering training/inference |
Common: description → term. E.g., "autonomously use tools and act toward a goal" = agent; "build agents" = Vertex AI Agent Builder; "base answers on trusted sources" = grounding; "ground recency via Google Search" = Grounding with Google Search; "specialized hardware powering training/inference" = AI infrastructure (TPU/GPU).
Watch the mix-ups: (1) agents (autonomous action) go a step beyond a chatbot that just returns text. (2) Grounding is central to mitigating hallucination, via Google Search or your own data (RAG). (3) Leaders need the "affects cost/sustainability" point, not TPU/GPU details.
2.3.3Section summary
- Agent (agentic AI) = autonomously use tools and act toward a goal; build with Vertex AI Agent Builder
- Grounding = base answers on trusted sources (Google Search / your data = RAG); mitigates hallucination
- AI infrastructure (TPU/GPU) powers large-scale AI; relates to cost/sustainability
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. What is AI that, beyond returning text, autonomously plans and uses tools to act toward a goal called?
Q2. What technique bases a model answer on trusted sources to reduce hallucination?
Q3. Which Google Cloud means is used to build agents and search apps?
Q4. Which grounding suits reflecting recent information in answers?
Q5. Which Google specialized hardware powers training/inference for large-scale generative AI?
Q6. To answer accurately based on internal documents, which approach is most appropriate?