What is Agentic AI?

Introduction

Agentic AI is software where one or more LLMs autonomously plan and execute multi-step tasks by calling tools, retrieving knowledge, reflecting on results and adjusting their plan — until the goal is achieved or the agent stops.

Beginner analogy: ChatGPT is a vending machine — one coin (prompt) = one snack (reply). An agent is a personal assistant — you give it a goal ("plan my trip to Tokyo under $2000") and it searches flights, compares hotels, checks your calendar and books — taking many small steps on its own.

Understanding the topic

Core concepts:

• Goal-directed: user gives a goal, not a single instruction.
• Plans: breaks goals into ordered/parallel subtasks.
• Uses tools: APIs, code execution, search, databases.
• Has memory: remembers earlier steps and past sessions.
• Reflects: critiques its own output and retries on failure.
• Autonomous: runs without a human in the loop for every step.

Syntax reference

Visual workflow / architecture:

User Request
      │
      ▼
┌──────────────┐
│   Planner    │  decomposes the task
└──────┬───────┘
       │ subtasks
       ▼
┌──────────────┐
│  Tool Caller │──► External APIs / DBs / Search
└──────┬───────┘
       │ observations
       ▼
┌──────────────┐
│  Reasoner    │  reflects · evaluates
└──────┬───────┘
       │ next step or "done"
       ▼
┌──────────────┐
│   Memory     │  short-term + long-term
└──────┬───────┘
       ▼
   Final Response

Real-world use

Devin (Cognition) plans and writes entire pull requests autonomously. Operator (OpenAI) controls a browser to book restaurants. Copilot Workspace turns an issue into a multi-file PR. Salesforce Agentforce handles customer cases end-to-end.

Best practices

• Start with a single tool and a single goal — expand only when stable.
• Always log every plan, tool call and observation for debugging.
• Set hard limits: max steps, max tokens, max cost per run.

Common mistakes

• Giving the agent 50 tools day one — it gets confused and loops.
• No max-step limit → infinite loops + huge bills.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What makes an AI system 'agentic'?
• Q2. Difference between an LLM call and an agent run?
• Q3. Name the four core capabilities of an agent.
• Q4. Why must agents have a max-step limit?
• Q5. Scenario: design an agent that triages GitHub issues. What tools does it need?