Welcome back to our Agentic AI with LangGraph documentation series. While our first post explored the high-level evolution from Generative AI to Agentic AI, this second installment focuses on a formal, detailed study of what defines an AI agent and the internal machinery that makes it work.
What is Agentic AI? #
Agentic AI is a software paradigm where a system is given a high-level goal or task by a user and works toward completing it independently with minimal human guidance. Unlike reactive systems that wait for individual prompts, an agent plans, takes action, adapts to changes, and seeks help only when necessary.
To visualize this, imagine planning a trip to Goa. A traditional Generative AI is reactive; you must ask it specific questions about flights, then hotels, then itineraries one by one. Conversely, an Agentic AI system is proactive and autonomous; you provide the dates, and it independently researches travel methods, recommends hotels, and builds the entire itinerary for you.
The Six Key Characteristics of an AI Agent #
To identify if a system is truly agentic, it must exhibit these six core traits:
1. Autonomy
Autonomy is the ability to make decisions and take actions to achieve a goal without needing step-by-step instructions. An agent is proactive, meaning it takes initiative before a human asks. This autonomy manifests in:
- Execution: Automatically moving through a multi-step plan.
- Decision Making: Independently choosing which candidates to shortlist based on criteria.
- Tool Usage: Deciding which tool (email, calendar, search) is needed at any given moment.
2. Goal-Oriented
An agent operates with a persistent objective in mind. Every action it takes is directed toward that specific target, which acts as a “compass” for its autonomy. These goals can be independent (e.g., “Hire a back-end engineer”) or constrained (e.g., “Hire only from India with a $1,000 budget”).
3. Planning
Planning is the agent’s ability to break a high-level goal into a structured sequence of sub-goals. This is often an iterative process:
- Candidate Plan Generation: The agent creates multiple possible paths to the goal.
- Evaluation: It assesses these plans based on efficiency, tool availability, cost, risk, and alignment with constraints.,
- Selection: It chooses the most optimized path to execute.
4. Reasoning
Reasoning is the cognitive process of interpreting information, drawing conclusions, and making decisions. An agent uses reasoning during planning (to decompose goals) and during execution (to handle errors or decide if human help is needed).,,
5. Adaptability
Adaptability allows an agent to modify its strategy in response to unexpected conditions while staying aligned with the original goal. If a tool fails (e.g., a LinkedIn API is down) or external feedback is negative (e.g., no one is applying to a job post), the agent finds an alternative path, such as messaging the user directly or suggesting a JD revision.,
6. Context Awareness
This is the ability to understand and retain relevant information from ongoing tasks, past interactions, and the environment. An agent must remember its original goal, current progress, and user preferences (e.g., “the user prefers interview questions in a Google Doc”) to function correctly over multiple days.,,
The Anatomy of an Agent: 5 Core Components #
To build an agent using frameworks like LangGraph, you need to understand these five high-level components:
| Component | Description |
|---|---|
| The Brain (LLM) | The core reasoning engine. It interprets goals, creates plans, selects tools, and generates natural language for communication., |
| The Orchestrator | The “manager” or nervous system. It executes the LLM’s plan step-by-step, handles conditional routing, retries failed steps, and manages task delegation., |
| Tools | The “hands and legs” of the agent. These are APIs, databases, or RAG-based knowledge bases that allow the agent to interact with the external world. |
| Memory | Stores information across two layers: Short-term (current session data like immediate tool results) and Long-term (high-level goals, user preferences, and guardrails). |
| Supervisor | Manages the Human-in-the-Loop (HITL) aspect. It enforces guardrails, requests human approval for high-risk actions, and escalates edge cases., |
Controlling Autonomy: Human-in-the-Loop #
Giving an agent full autonomy can be risky (e.g., it might spend too much on ads or send incorrect offer letters). To mitigate this, developers implement controls:
- Scope Definition: Limiting the tools or actions an agent can perform independently.
- Checkpoints: Requiring human approval before critical steps (like posting a job).
- Override Controls: Allowing users to pause, stop, or change the agent’s behavior at any time.
- Guardrails: Hard rules the agent must follow, such as “never schedule interviews on weekends.
Q.1 Why is an Orchestrator considered one of the most important components of an Agentic AI system?
It permanently stores all user conversations for future training.
It coordinates planning, tool execution, memory usage, and workflow until the goal is achieved.
It replaces the need for a Large Language Model.
It converts all documents into embeddings before training.
Explanation
The Orchestrator acts as the central coordinator of the agent. It manages task execution, invokes tools, tracks state and memory, handles retries, and ensures all components work together to accomplish the user’s objective efficiently.
Q.2 In the context of Agentic AI, what does 'Autonomy' specifically refer to?
The requirement for constant human supervision for every sub-task.
The ability to function without any Large Language Model.
The ability to make decisions and take actions independently to achieve a goal.
The speed at which the AI generates text responses.
Explanation
Autonomy means the agent can reason, select tools, make decisions, and execute tasks with minimal human intervention while working toward a specified goal.
Q.3 Why would an Agentic AI system generate multiple 'candidate plans' during the planning stage?
Because the LLM is unable to decide on a single sequence of actions.
To ensure the human user has more work selecting one.
To treat planning as a search problem to find the most optimized path.
To use more computational resources to improve accuracy.
Explanation
Advanced agents often generate multiple possible plans and evaluate them before execution, selecting the one most likely to achieve the objective efficiently and reliably.
Q.4 Which component of an Agentic AI system is responsible for the actual step-by-step execution and sequence management of a plan?
The Tools.
The Memory.
The Brain.
The Orchestrator.
Explanation
The Orchestrator coordinates workflow execution, manages task sequencing, invokes tools, tracks progress, and determines what should happen next during the agent’s execution.
Q.5 How does an Agentic AI system demonstrate 'Adaptability' when a specific tool fails, such as a Calendar API being down?
It automatically rewrites the Calendar API source code.
It stops all operations until the API becomes available.
It modifies its plan and seeks an alternative approach, such as asking the user or using another tool.
It ignores the failure and continues without the missing information.
Explanation
Adaptability enables an AI agent to recognize failures, revise its plan, choose alternative tools or strategies, and continue progressing toward the original objective.
Q.6 In the context of Agentic AI memory, what is typically stored in 'Short-term Memory'?
Universal ethical guardrails and company-wide hiring policies.
The original source code of the Large Language Model.
The history of the current conversation and recent tool outputs.
User preferences that never change across different years.
Explanation
Short-term memory stores temporary information such as recent messages, intermediate reasoning, retrieved documents, and tool outputs that are relevant to the current session.
Q.7 What is the role of the 'Supervisor' component in an Agentic AI architecture?
To generate candidate plans using the LLM.
To provide external access to databases and search engines.
To store long-term user preferences.
To manage Human-in-the-Loop interactions and enforce guardrails.
Explanation
The Supervisor monitors the agent’s behaviour, applies safety policies, requests human approval when required, and ensures actions remain within predefined guardrails.
Q.8 Which component of an Agentic AI system is primarily responsible for deciding whether to retrieve information from a Knowledge Base, call an external API, or ask the user for clarification?
Memory
The LLM (Brain)
The Embedding Model
The Vector Database
Explanation
The LLM, often referred to as the Brain of the agent, performs reasoning based on the user’s goal and the current context. It decides whether to retrieve information using RAG, invoke an external tool or API, or request additional clarification from the user before proceeding.
Q.9 In an Agentic AI system, how does a Knowledge Base (using RAG) typically fit into the high-level components?
It functions as the Supervisor for human interactions.
It replaces the need for an Orchestrator.
It is considered a type of Tool.
It serves as the system's Brain.
Explanation
A RAG-based Knowledge Base is typically exposed as a retrieval tool that the agent can invoke whenever it needs domain-specific information from external documents.
Q.10 Which characteristic is described as a 'Compass' that directs all of an Agentic AI's autonomous actions?
Goal-Orientedness.
Context Awareness.
Adaptability.
Reasoning.
Explanation
Goal-Orientedness ensures that every decision, plan, and action taken by the agent is aligned with achieving the user’s desired outcome rather than simply responding to individual prompts.
