What's changed: Added per-section figures (cert-figure-retrofit). New AI-901 Chapter 4 (Domain 2: agent definition = goal→autonomous plan→tool execution→loop, building blocks = instructions/model/knowledge/tools, function calling, difference from generative AI, Foundry Agent Service, multi-agent, safety = least privilege/human-in-the-loop/Content Safety)
4.1What Is an AI Agent
Understand that agentic AI extends generative AI with autonomous goal-directed planning, tool use, and multi-step execution; learn how it differs from simple generative AI, the building blocks (tools/function calling, knowledge, actions), and when agents fit.
An AI agent is a "reason-then-act" AI that, given a goal, autonomously plans steps, calls the tools it needs, executes, observes the result, and decides the next move. Whereas simple generative AI "returns one piece of text per input," an agent loops through multiple steps until the goal is met—the decisive difference. For "arrange next week’s business trip," an agent autonomously checks availability, books, and sends a confirmation email using tools.
4.1.1The building blocks of an agent
- Instructions: the agent’s purpose, role, and constraints (akin to a system message).
- Model: the LLM that does the reasoning and text generation (deployed in Foundry).
- Knowledge: the information sources it references (e.g., internal docs); supplied via grounding/RAG.
- Tools / actions: the external functions the agent can call (function calling, APIs, search); the ability to act is the essence of an agent.
4.1.2How tools (function calling) work
An agent can "act" thanks to function calling (tool calling). Developers register the agent’s available "tools (functions)" and how to use them. When the agent judges a tool is needed to reach the goal, it decides when and with what arguments to call it, then proceeds based on the result. This lets the AI chain real-world operations—inventory lookup, booking, sending email, calculation.
| Aspect | Simple generative AI | Agent |
|---|---|---|
| Behavior | One input → one response | Goal → autonomously loop multiple steps |
| External actions | None (just returns text) | Executes via tools/function calling |
| Best for | Summary, drafting, classification | Multi-step tasks: arrange, research-then-act |
Scenario: a support agent. Goal "process the customer’s return request" → (1) check the order DB via a search tool → (2) decide eligibility → (3) if eligible, run the return-registration tool → (4) notify the customer via the email tool → (5) log a summary. It autonomously chains multiple tools to complete the task. Simple generative AI could only draft the reply text.
Watch out: (1) do not confuse generative AI (returns text) with an agent (acts via tools, multi-step)—if there is "action / multiple steps," it is an agent. (2) Because agents act autonomously, wrong actions cause real harm → least privilege and human approval for critical actions (human-in-the-loop). (3) Tools (function calling) let the AI call external functions; it is not retraining the model.
Common: (1) "given a goal, autonomously executes steps and external actions" = agent. (2) "mechanism for the AI to call external functions (API/search/booking)" = tools / function calling. (3) "only summarize or draft" = simple generative AI suffices (no agent needed). (4) "reduce autonomous-execution risk" = least privilege, human-in-the-loop.
4.1.3Section summary
- Agent = a "reason-then-act" AI that loops goal → plan → execute tools → act on results
- Building blocks: instructions / model / knowledge (grounding) / tools & actions (function calling)
- Difference from simple generative AI = whether there are external actions and multiple steps
- Autonomous-execution risk → keep it safe via least privilege and human-in-the-loop
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Which AI, given a goal, autonomously plans and executes multiple steps using tools?
Q2. What mechanism lets an agent "act" by calling external functions (search, booking, email)?
Q3. What is the most fundamental difference between simple generative AI and an agent?
Q4. Which building block lets an agent base decisions on trustworthy info such as internal docs?
Q5. Which is best for the task "summarize the body of a single email"?
Keep track of your progress
The full study guide is free to read. Sign up free to practice with the question bank, track what you have read, review your mistakes, and highlight passages.