While Large Language Models (LLMs) excel at generating text, answering questions, or summarizing information based on a single prompt, their ability to address complex problems expands significantly when they operate within an agentic workflow. But what exactly does this term mean, and how does it differ from simply sending a series of requests to an LLM?

At its core, a workflow is a sequence of steps or operations designed to achieve a specific outcome. You encounter workflows everywhere, from a manufacturing assembly line to the process of publishing a blog post. In the context of AI, an agentic workflow elevates this by incorporating autonomy and intelligence, primarily driven by an LLM.

An agentic workflow is a system where an AI agent, often powered by an LLM, autonomously plans, executes, and adapts a series of actions to achieve a predefined goal. Unlike a simple script that executes a fixed set of instructions, an agent in such a workflow can typically:

Decompose Goals: Break down a high-level objective (e.g., "plan a weekend trip to San Francisco") into smaller, manageable sub-tasks (e.g., find flights, book accommodation, list attractions).
Make Decisions: Choose the next best action based on the current state, available tools, and the overall goal. This might involve deciding which tool to use, what parameters to pass to it, or how to interpret its output.
Utilize Tools: Interact with external systems, APIs, or data sources to gather information or perform actions outside its inherent capabilities (e.g., searching the web, accessing a database, running code).
Maintain Context: Remember previous steps, observations, and decisions to inform future actions, often managing a form of memory.
Iterate and Adapt: Potentially adjust its plan or retry steps if an action fails or if new information becomes available.

Consider the difference. A standard LLM interaction might involve you asking, "What's the weather in Paris?" The LLM responds directly. In an agentic workflow, you might ask an AI travel agent, "Find me the best way to get from London to Paris for a meeting next Tuesday morning, arriving before 9 AM, and book the ticket." The agent would then initiate a workflow:

Step 1 (Planning & Reasoning): Identify travel parameters (date, time, origin, destination, constraints). Decide to check both flights and trains.
Step 2 (Tool Use - Query): Check flight APIs for London to Paris on the specified date.
Step 3 (Tool Use - Query): Check train APIs (e.g., Eurostar) for the same.
Step 4 (Reasoning & Decision): Compare flight and train options based on arrival time, duration, and perhaps stored user preferences for cost or comfort.
Step 5 (Interaction/Clarification - Optional): If multiple suitable options exist, it might ask for clarification: "I found a flight arriving at 8:30 AM for $150 and a train arriving at 8:45 AM for$ 120. Which do you prefer?"
Step 6 (Tool Use - Action): Proceed to book the chosen option using the relevant API.
Step 7 (Confirmation): Report back the booking details.

This multi-step, decision-driven process, often involving external interactions, is characteristic of an agentic workflow. It's a significant advancement from single-shot LLM queries. Traditional automation or scripting can also perform multi-step tasks, but they typically rely on explicitly programmed logic for every decision point and every step. If an unexpected situation arises, a traditional script often fails or requires manual intervention.

Agentic workflows, by contrast, use the LLM's reasoning capabilities to navigate these complexities. The "program" is less about explicit step-by-step instructions and more about defining the goal, the available tools, and the general strategy the agent should employ. Your prompts become the primary interface for shaping this strategic behavior, a topic we'll explore extensively throughout this course.

The ability to construct these workflows allows AI to address tasks that are too multifaceted or dynamic for simple LLM prompting alone, opening doors to more sophisticated automation and problem-solving.

A comparison: a direct request-response model versus a multi-step, decision-driven agentic workflow.

The main point is that agentic workflows enable AI to move from being a sophisticated text generator or question answerer to a more autonomous problem-solver. This shift requires a different approach to how we interact with and guide these systems, particularly through careful prompt engineering. As we proceed, you'll learn how to craft prompts that define these workflows, guide agent decision-making, and ultimately build more capable AI applications.

Was this section helpful?