AI Agents in the Wild

1. AI Agents in the Wild

Welcome to the course!

2. Meet your instructor

I'm Korey Stegared-Pace, a Senior AI Cloud Advocate at Microsoft, and I'll be your instructor for this course.

3. The journey ahead!

In this course, you'll learn about what it takes to design and scale AI agents successfully. We'll start by looking at what successful AI agents have in common, and their typical failure points.

4. The journey ahead!

We'll learn about frameworks for integrating agents with surrounding systems, including the powerful Model Context Protocol, or MCP, developed by Anthropic, and Agent-to-Agent, or A2A, developed by Google. These interoperability frameworks will be key in scaling and maintaining our agents.

5. The journey ahead!

Finally, problems are inevitable with AI agents, so we'll discover how to set these systems up for success in production and fail gracefully. To begin, let's remind ourselves about how AI agents work.

6. Components of an agent

When a user sends a prompt to an agent,

7. Components of an agent

it enters a large language model, or LLM, which interprets the task and decides what steps to take.

8. Components of an agent

To complete the task, the model may decide to request the use of a tool, which are functions to interact with the external world. These tools could retrieve up-to-date information from databases or APIs, run code, or trigger events. These tool results are passed back to the LLM to be used in its response.

9. Components of an agent

This workflow of integrating models and tool calls is managed by an orchestration layer, which often handles things like memory and logging.

10. An agentic application

When you interact with an AI agent application, the "agentic" part is just one piece of the puzzle.

11. An agentic application

Users commonly interact with agents through a conversational interface, which is often a free text box, but could include dropdowns, selectors, and other elements for modifying parameters. The message sent by the user is often combined with additional "system" messages to craft a prompt to optimize the model outputs. Zooming-in on the model - this is usually hosted on a server, separate from the application. For many use cases, the model is hosted by the model provider, like those shown, but they can be self-hosted. Lastly, data storage might be used by the tools or for collecting user data, so it might be integrated with several other components. Throughout the course, we'll discuss how to get these components working optimally and how to mitigate common failure points.

12. When to use agents?

While there is much excitement to apply agents in a variety of ways, they aren't necessarily the solution to every problem. Open-ended problems are difficult to create deterministic workflows for due to the number of possible options, so these represent a strong opportunity for agentic automation. Other problems require multiple steps and actions to complete them,

13. When to use agents?

and this is where AI Agents really shine. They can plan out the steps required to solve a complex problem, and make the necessary tool calls to complete them one-by-one.

14. When to use agents?

Finally, AI agents thrive in changing circumstances. Some situations rapidly evolve, or user preferences get updated, and agents can use this data to respond differently. Let's look at two scenarios.

15. Example: sending reminder emails

An online training provider wants to automate sending reminder emails to participants that haven't RSVP'd to the email invite. This scenario

16. Example: sending reminder emails

isn't open-ended, as all participants will either be sent a reminder or not,

17. Example: sending reminder emails

nor is it complex, as there's two user groups corresponding to two outcomes,

18. Example: sending reminder emails

and it isn't required to be adaptable as the input information - the user list - isn't evolving. This problem likely has an IT solution rather than an agentic solution.

19. Example: IT support agent

In the same company, the IT department are receiving lots of questions about the products and software available for employees, which is impacting their ability to deliver important rollouts. There are lots of ways that employees may ask about these products or software, and this list of products may even change over time.

20. Example: IT support agent

This open-ended and complex problem requires an adaptable solution to respond to these user questions. AI agents are worth experimenting with here.

21. Let's practice!

Let's put these fundamentals into practice!

Create Your Free Account

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.