Agentic AI & ReAct

Up until now, we have built Pipelines.

A pipeline (like RAG) is a hardcoded sequence of events:

1. User asks a question.
2. Embed the question.
3. Search the database.
4. Generate an answer.

Pipelines are deterministic and highly reliable, but they are rigid. If the user asks a question that requires searching the database twice, or requires searching the database and then using a calculator, a standard RAG pipeline will fail. It only knows how to do exactly what you programmed it to do.

What is an AI Agent?

An AI Agent is a system where the LLM acts as the "brain" or the routing engine. Instead of hardcoding the sequence of steps, you give the LLM a goal and a set of tools, and you let the LLM decide which tools to use and in what order.

• Perception: The agent receives input from the user or its environment.
• Cognition: The agent reasons about the input, breaks down the problem, and decides on a plan.
• Action: The agent uses tools (Function Calling) to execute the plan.

The Shift in Control

When you build an agent, you are surrendering control flow to the LLM.

In a pipeline, your Python or JavaScript code says: if X, then Y. In an agentic system, your code says: while goal_is_not_met: ask_llm_what_to_do_next().

This shift allows agents to handle incredibly complex, multi-step, and ambiguous requests that would be impossible to hardcode. However, it also introduces unpredictability. The agent might get stuck in an infinite loop, use the wrong tool, or hallucinate a conclusion.

To make agents reliable, we need a structured framework for how they "think." The most famous of these frameworks is ReAct.