Agentic AI: Workflows vs. Agents

Video Introduction

This video is the first episode of Season 2 of "Real Terms with AI," which clearly defines the concepts of AI agents and agentic workflows, covering their differences and use cases. Hosts Aja Hammerly and Jason Davenport explain AI-related topics using terms software developers can actually understand, with this season focusing specifically on agentic AI.

Core Concepts: What Are Agents and Agentic Workflows?

1. What is an agent?

   • Jason: "An agent is a system that performs tasks in a nondeterministic manner. That is, even with the same input, it may not always produce the same result."
   • When an agent is given a goal, it defines and executes the tasks needed to achieve that goal on its own, then reviews (reflects on) the results to verify they meet the objective.
   • Example: "If you ask an agent to 'build me a website with a snake game,' the agent will plan the necessary tasks on its own, execute them, and review the results to achieve the goal."

2. What is an agentic workflow?

   • Aja: "A workflow operates in a deterministic manner, producing predictable results. It clearly defines specific tasks and outcomes and executes the process accordingly."
   • Example: "A workflow can be illustrated by an invoice processing process. An LLM is used to analyze invoices and extract data, which is then entered into a form. But ultimately the user reviews the data and decides whether to upload it."

3. Agents vs. Workflows: Key Differences

   • Agents perform tasks autonomously, with no fixed order or method.
   • Workflows operate in a more structured and predictable way, requiring more human intervention.
   • Aja: "In many cases, the problems we're trying to solve can be more appropriately addressed with an agentic workflow rather than an agent."

Use Cases and Examples

1. Workflow Example

   • Invoice processing: "An LLM is used to analyze invoices and extract data, which is then entered into a form. But ultimately the user reviews the data and decides whether to upload it." A deterministic and predictable process.

2. Agent Example

   • Building a snake game website: "When the agent receives the goal 'build me a website with a snake game,' it defines the necessary tasks on its own, executes them, and reviews the results to achieve the goal." An autonomous and nondeterministic approach.

Why Does This Matter?

• Jason: "Understanding agentic behavior is closely connected to human interaction. For example, it's similar to thinking about who can make what decisions in a business environment - how to distribute decision-making authority (agency)."
• Aja: "Whether to use an agent or a workflow depends on the nature of the problem. For example, when building a customer service agent, you need to consider the possibility that the agent might make decisions a human wouldn't."

Considerations When Choosing Between Agents and Workflows

1. When an agentic workflow is appropriate

   • When you want to add LLMs to existing deterministic processes to increase efficiency.
   • Examples: data summarization, simple information processing, etc.
   • Aja: "In these cases, it's important to design the process so that humans can intervene to review and adjust results."

2. When an agent is appropriate

   • When you can clearly define goals and outcomes, and autonomy is needed.
   • Examples: complex problem solving, creative tasks, etc.
   • Aja: "Agents are useful in situations where they can autonomously perform tasks to achieve goals."

Tips and Resources for Viewers

• Jason: "This season, we'll share various code snippets and architecture examples on GitHub. This will help you concretely implement abstract concepts."
• Aja: "Check out the examples of agents and agentic workflows through the links and try applying them to your projects!"

Wrap-Up

• Aja: "This season, we'll deeply explore the concepts of agents and agentic workflows and learn how to utilize them. Let's learn together!"
• Jason: "Happy prompting!"
• Aja: "Happy prompting!"

Keyword Summary

• Agent: Autonomous, nondeterministic, goal-oriented.
• Agentic workflow: Deterministic, predictable, human intervention.
• LLM (Large Language Model): Data processing and analysis tool.
• Agency: Distribution of autonomy and responsibility.

This video helps you clearly understand the difference between agents and workflows so you can choose the right approach for your projects