Agentic AI

Introduction

Agentic AI is the final and most advanced topic on the exam — it builds on everything else (models, prompting, ethics, data, development) and asks: what changes when an AI system doesn't just respond to a prompt, but plans and takes autonomous action toward a goal? This topic covers what makes a system "agentic" rather than purely generative, how agents are designed and orchestrated, the Model Context Protocol (MCP) that standardizes how agents connect to tools and data, human-in-the-loop strategies that keep autonomy safe, and the unglamorous-but-critical data transformation work that holds multi-step agent pipelines together. Expect scenario questions that ask you to judge whether something is "truly" agentic, spot a missing safeguard, or pick the right level of human oversight.

Key Concepts

Agentic AI — AI systems that can plan and autonomously execute a sequence of actions toward a goal, as opposed to Generative AI, which primarily produces content in response to a single prompt.

Autonomy — The ability of a system to decide its own next action without a human prompting each step.

Agent design principles — Scoping an agent's tools narrowly (least privilege), defining clear goals, building in retry limits/escalation paths to prevent runaway loops, and logging reasoning for auditability.

Orchestration — Coordinating multiple agents (or steps) toward an overall goal, often via a central orchestrator delegating to specialized worker agents.

Model Context Protocol (MCP) — A standardized protocol for connecting AI applications to external tools, data, and context, avoiding one-off custom integrations per model-to-tool pairing. Key primitives: resources (read-only context/data), tools (actions the model can invoke), and prompts (reusable interaction templates). MCP servers expose these; MCP clients (the AI application) connect to and use them.

Human-in-the-loop (HITL) — A design pattern where a human reviews or approves certain agent decisions before they take effect, typically reserved for high-risk or hard-to-reverse actions, using strategies like confidence thresholds or risk-based escalation.

Data transformation and mapping — Converting and matching data between systems with different formats or schemas (e.g., field names, date formats, numeric types) so one tool's output can be correctly consumed by the next step in an agent pipeline.

How It Works

1. A task is defined with a goal. The agent (or orchestrator, in a multi-agent system) plans a sequence of steps to reach it, rather than executing a single fixed script.
2. The agent uses tools — scoped narrowly to what the task actually requires — often connected via MCP, which standardizes how those tools and any supporting context (resources) are exposed and consumed.
3. As the agent moves between steps, data frequently needs to be transformed or mapped — converting formats, reconciling schema differences — before it's valid input for the next tool. Validation at each step boundary catches errors before they compound.
4. Higher-risk or harder-to-reverse actions are routed through a human-in-the-loop checkpoint, calibrated by risk and reversibility (or a confidence score), rather than either requiring review of everything (too slow) or nothing (too risky).
5. Throughout, safeguards constrain autonomy: retry limits prevent runaway loops, logging supports auditability, and tool scoping limits the blast radius if something goes wrong — autonomy is powerful but needs the same engineering discipline as any other production system, not less.

Commands / Syntax / Key Values

• No specific commands for this topic — it's architecture and design judgment, not syntax. Key terms to know cold: the Agentic vs. Generative distinction, the three MCP primitives (resources/tools/prompts), and risk-based HITL routing.

⚠ Exam Traps

• Calling anything without human input "agentic" — a rigid hardcoded script with no human input is not agentic; agency requires the system itself to reason about and choose among actions, not just execute a fixed sequence.
• Treating agency as binary — the exam favors describing it as a spectrum based on scope of autonomous decision-making, not a strict yes/no label.
• Assuming a human sign-off always means genuine oversight — rubber-stamped approvals (no time, no context, no real authority to override) are procedural compliance, not meaningful HITL.
• Confusing MCP resources and tools — resources are read-only context; tools let the model take an action that changes state. Know the difference.
• Treating data mapping as a trivial/invisible step — silent field drops, missing validation, or undetected schema mismatches are common exam scenarios testing whether you catch the risk.
• Assuming more autonomy is always better — the exam rewards recognizing when added autonomy (self-modifying prompts, uncapped sub-agent spawning) introduces risk without corresponding benefit.

Practice Questions

Q1. A system follows a fixed if/else script with no model-driven decision-making, but never asks for human input during execution. Is it agentic? - A. No — absence of human input alone doesn't make something agentic - B. Yes — any system without human input during execution is agentic - C. Yes — fixed scripts are advanced agentic reasoning - D. No — because it doesn't use a database

Answer: A — Agency requires the system to reason about and choose its own actions, not just execute a hardcoded sequence.

Q2. What is the functional difference between an MCP "resource" and an MCP "tool"? - A. They are functionally identical - B. A resource provides read-only context/data; a tool lets the model take an action that changes state - C. A tool is only used for billing - D. A resource is always faster than a tool

Answer: B — Resources are for context; tools are for action.

Q3. A HITL reviewer approves every single agent recommendation without scrutiny because "the agent is usually right." What does this represent? - A. Rubber-stamping — the checkpoint no longer provides genuine independent review - B. Proof the system is working perfectly - C. A reason to give the agent more autonomy - D. A reason to remove the reviewer

Answer: A — A checkpoint that's no longer scrutinized provides only the appearance of oversight, not the substance.

Summary

• Agentic AI is defined by autonomous planning and action-taking toward a goal — not merely the absence of human input, and not a binary label but a spectrum.
• MCP standardizes how agents connect to tools, resources, and prompts, avoiding custom one-off integrations — know the three primitives and that externally-sourced content should be treated as untrusted.
• Human-in-the-loop should be risk-based and genuinely substantive (real authority, time, and context to override), not a rubber stamp — and data transformation/mapping between pipeline steps needs validation and auditability, since silent errors compound quietly across multi-step agent workflows.