Skip to main content

Lesson 7: Generative vs. Agentic AI

Topics Covered
  • The Paradigm Shift: From generating artifacts to delivering outcomes.
  • Generative AI: The linear "Prompt-Response" cycle.
  • Agentic AI: The circular "Perceive-Decide-Act" loop.
  • The LLM's Role: Shifting from the "Mouth" to the "Brain."
  • Risk Profile: The cost of autonomy in production systems.

For the first few years of the AI boom, we focused on Generative AI (ChatGPT writing poems). Now, we are shifting to Agentic AI (software doing work). The difference isn't just capability; it's a fundamental change in architecture.

1. Generative AI: The Creator (Linear)​

Generative AI is reactive. It waits for a human to push a button, and it produces a digital artifact (text, code, image).

  • The Flow: Input -> Model -> Output.
  • The Goal: Create a high-quality draft.
  • The Human's Role: The human must take that output and do something with it (e.g., copy-paste the code, send the email).
  • Example: You ask AI to write a SQL query. It gives you the text. You must copy it into your database tool to run it.

2. Agentic AI: The Doer (Circular)​

Agentic AI is proactive. It accepts a high-level goal and works autonomously to achieve an outcome.

  • The Flow: Goal -> [Perceive -> Reason -> Act -> Observe] (Loop) -> Outcome.
  • The Goal: Solve a problem.
  • The Human's Role: Define the boundary and approve the final result.
  • Example: You ask an AI Agent to "Get me the sales data." It writes the SQL query, connects to the database tool, executes the query, fixes any syntax errors it encounters, and returns the final CSV file.

3. The Architecture Shift​

FeatureGenerative AIAgentic AI
LLM RoleThe Mouth. Its job is to speak/write fluently.The Brain. Its job is to reason and plan steps.
Key DependencyTraining Data. It needs to "know" facts.Tools (APIs). It needs "hands" to interact with software.
Success MetricCoherence & Style.Success Rate & Accuracy.
InteractionTurn-based chat.Autonomous loops.

4. How They Work Together​

Both systems use the same underlying models (like GPT-4 or Claude), but they use them differently.

In an Agent, the LLM uses its "generative" capability to create an Internal Monologue (Chain of Thought).

  • Thought: "The user wants sales data. I should check the sales_db."
  • Action: Call get_tables() tool.
  • Observation: "Table found: monthly_revenue."
  • Thought: "Now I will query that table."

The model "talks to itself" to reason through the problem, rather than talking to the user.

5. The Risk Profile (Why this matters)​

Understanding the difference is critical for safety.

  • Generative Risk: The model might write something offensive or incorrect (Hallucination). Cost: Low (User ignores it).
  • Agentic Risk: The model might execute a wrong command (e.g., "Delete Table" instead of "Select Table"). Cost: High (Data loss or financial impact).

Building Agentic AI requires Guardrails—code that sits between the Agent and the Tool to prevent dangerous actions before they happen.