What is Agentic AI? | Agentic AI using LangGraph | Video 2 | CampusX

Summary

The video introduces Agentic AI as an AI system that receives a goal and works toward completing it with limited step-by-step human guidance. Unlike a typical generative-AI chatbot, which responds to individual prompts, an agentic system can plan, take actions, monitor progress, and adjust its approach. The lesson provides conceptual groundwork for later videos about building agents with LangGraph.

Agentic AI Compared with a Reactive Chatbot

The presenter uses planning a trip to Goa to illustrate the difference:

• With a reactive chatbot, the user asks separate questions—for example, about transport, hotels, or places to visit—and the chatbot answers each one.
• With an agentic system, the user gives a broader goal, such as planning a trip for specific dates. The system can work through related tasks and present a plan, while asking for human input when needed.

The distinction is that an agentic system is proactive and goal-directed, rather than simply responding to isolated prompts.

Example: An AI Recruiter

A central example follows an AI assistant tasked with hiring a remote backend engineer with two to four years of experience. Its work could proceed as follows:

1. Understand the goal and make a plan. Identify the hiring objective and outline stages such as drafting a job description, advertising the role, screening applicants, arranging interviews, and onboarding a hire.
2. Draft the job description. Consult company documents for relevant technologies, responsibilities, and other details, then ask the recruiter to review the draft.
3. Post and monitor the role. With permission, use available platform tools or APIs to post the role and track applications.
4. Adapt if results are weak. If too few people apply, suggest changing the job description or promoting the listing, then seek approval before proceeding.
5. Screen candidates. Use a résumé-parsing tool to assess applicants and report which candidates appear to be strong, partial, or weak matches.
6. Arrange interviews. Check the recruiter’s calendar, propose interview times, send or prepare invitations, and provide interview questions and reminders.
7. Prepare an offer and onboarding. After the recruiter selects a candidate, draft an offer letter for review, send it once approved, monitor the candidate’s response, and help initiate onboarding tasks.

The example illustrates autonomy while also showing that human review and authorization may remain important for sensitive or consequential actions.

Six Key Characteristics of Agentic AI

1. Autonomy

The system can make decisions and take actions toward a goal without needing instructions at every step. Autonomy can apply to executing tasks, making decisions, and choosing which tools to use.

The presenter emphasizes that autonomy should be controlled. Possible safeguards include:

• Limiting which tools or actions the agent can perform independently.
• Requiring human approval at selected checkpoints.
• Allowing people to pause, stop, or change the agent’s activity.
• Defining rules and ethical boundaries that the agent must follow.

These controls help reduce risks such as inappropriate hiring decisions, biased screening, unauthorized offers, or unapproved advertising spend.

2. Goal Orientation

The agent keeps a persistent objective in view and directs its actions toward achieving it, rather than merely reacting to individual prompts. Goals may include constraints—for example, hiring remotely, requiring particular experience, or staying within a budget.

Goals and their progress can be represented in memory, including status, completed tasks, and remaining work. A goal may also change mid-process, requiring the agent to revise its plan.

3. Planning

Planning turns a high-level goal into an organized sequence of actions and smaller subgoals. The process described is:

1. Generate candidate plans: Create multiple possible routes to the desired outcome.
2. Evaluate the plans: Compare them using factors such as speed, available tools, cost, risk, and fit with the user’s constraints.
3. Select a plan: Choose the best option, potentially using a predefined policy or human input.

Planning and execution can form an iterative loop: if a step fails or circumstances change, the agent can return to planning and choose a revised approach.

4. Reasoning

Reasoning enables the agent to interpret information, draw conclusions, and make decisions. It is needed both when planning and when carrying out a plan. Examples include:

• Breaking a goal into tasks.
• Selecting a tool.
• Estimating dependencies or risks.
• Choosing between candidate actions.
• Deciding when to ask a person for help.
• Responding to errors.

5. Adaptability

The system can revise its plans, strategies, or actions in response to unexpected conditions while still pursuing the goal. Triggers may include an unavailable tool, feedback from the environment, poor results, or a changed goal.

For example, if a calendar service is down, the agent might ask the recruiter directly about availability. If a job listing receives few applications, it might suggest revising the listing or promoting it.

6. Context Awareness

The agent needs to retain and use relevant information across a multi-step task. Context may include the original goal, progress so far, conversations with the user, the state of the environment, tool results, user preferences, and applicable rules.

The presenter describes two broad types of memory:

• Short-term memory: Information relevant to the current session or task, such as recent tool results.
• Long-term memory: Persistent information such as user preferences, past interactions, or standing rules.

Memory helps prevent an agent from losing track of prior decisions and task progress.

Five High-Level Components of an Agentic AI System

1. Brain: In the LLM-based systems discussed, the language model interprets the user’s goal, supports planning and reasoning, selects tools, and handles communication.
2. Orchestrator: Coordinates the plan’s execution. It manages task order, conditional routing, retries, loops, and whether work should be handled by the model or delegated to a person. The presenter notes that frameworks such as LangGraph can be used to build this kind of orchestration.
3. Tools: Let the agent interact with external systems—for example, APIs, calendars, email, databases, résumé parsers, and company knowledge sources.
4. Memory: Stores task context, goals, progress, tool responses, prior interactions, preferences, and other information needed across steps or sessions.
5. Supervisor: Supports human oversight, including approvals for risky actions, enforcement of guardrails, and escalation of unusual cases.

The video closes by framing these ideas as foundational concepts: define what an agent is, understand its key traits, and recognize the components that later practical implementations will use.

Speakers and Sources Featured

• Nitesh, the presenter and narrator, identified in the introduction as the host of the CampusX YouTube channel.
• No additional speakers or external sources are clearly featured in the subtitles.