By Jahanzeb Ahmad
It seems like conversations and discussions about AI are everywhere right now, and it is understandable to feel overwhelmed by it all. We hear many positive things about what AI can do, how everything has changed, and how this is the future. At the same time, there are valid concerns about how AI is being used. We also hear a lot of buzzwords and terminology we may not be totally familiar with. One moment the focus is on ChatGPT, then it is Gemini, then Claude. Sometimes the conversation is about AI chatbots, other times it is about AI agents. With so many terms, tools, and opinions circulating at once, it is easy to feel uncertain about where to focus.
As staff and faculty, we are often left asking practical questions. How can I use AI to actually help me do my job? And honestly, how am I even supposed to get started?
In this post, I want to address some of these questions while keeping things as simple as possible. Rather than surveying every tool available, we will focus on just one that many of us already know, ChatGPT. More specifically, we will look at how you can use it to create your first AI agent.
The good news is that getting started is really not as difficult as it seems, so let us get started.
What is an “AI agent”, anyway?
Let’s start with the basics. What do we mean by an “AI agent”, or just “Agent”? In this context, those terms are pretty much interchangeable.
An AI agent is a goal oriented digital assistant designed to help you complete a specific objective. Typically, AI agents are created using a large language model, such as OpenAI’s ChatGPT. The key difference is that an AI agent is given a persona and can make autonomous choices. An agent can also reflect on its responses, evaluate them, and adapt in order to improve over time with testing.
I find it helpful to think of an AI agent almost like a person, similar to an assistant helping you accomplish a specific goal. That also means that your assistant can only be as good as your prompt.
There is often confusion about agents, especially when they are described as “chatbots.” We tend to lump agents and chatbots together, but most chatbots are purely conversational and rely on pre scripted replies, like the clunky ones you have probably encountered on some websites. Chatbots cannot really assist you in completing a goal. They can only respond within a narrow conversational scope. Because of this, there is a clear distinction between a chatbot and an AI agent.
AI agents are typically built on large language models, and one of the more well known ways to create one right now is through ChatGPT. With the right prompting and structure, you can turn it into a surprisingly capable assistant.
Custom GPTs: An Entry Point into Designing AI Agents
As I mentioned earlier in this blog, this will be focused on OpenAI’s ChatGPT, although the thought process behind building and using agents applies to any Large Language Model you want to work with. It is important to note that a regular ChatGPT is already an agent, as it is able to complete tasks based on your prompts and feedback. So if you have been using ChatGPT, congratulations, you have already interacted with an agent.
The question is, how can you take this to the next level? Sure, you can start a chat with ChatGPT, have a simple conversation, and give it prompts to complete tasks, and it would work to some degree. However, there is a better way to create agents in ChatGPT.
With the paid version of ChatGPT, or a ChatGPT Edu license, you can create agents using a feature called Custom GPTs. This differs from a regular conversational chat with ChatGPT because you can give the agent a dedicated set of instructions that it follows every time you interact with it. Essentially, you build the prompt into ChatGPT so it already understands the context of the task you want it to complete.
Through this prompt, you can give the agent a persona, a specific goal, and instructions to achieve that goal, and provide relevant context files or links for it to reference. Essentially, you are giving the Custom GPT all the context it needs to consistently perform a specific task.
While in build mode, you can also interact with your agent, test how it responds, and make updates as you go. This makes it easy to refine and improve the agent over time. Overall, it is a practical and accessible way to get started with creating agents, with no advanced setup required.
See below for the steps to create a custom GPT within a Paid or Edu ChatGPT account.
In your paid or EDU ChatGPT account:
Click Explore GPTs in the left hand menu under GPTs.
Click the Create button.
You will now enter the custom GPT builder. Under the Create tab, you can begin by having a conversation with ChatGPT to describe what you want the agent to do.
In the Configure tab, you will be able to see the prompt and customize it in any way you like. You can also add files and adjust other settings.
On the right hand side, you can preview and test the customGPT agent in real time as you make changes.
The hard part: getting started and letting go of the need for perfection
Now we understand the basics of how to access the tool for creating an agent using ChatGPT’s CustomGPT builder. However, it’s only natural to feel stuck when it comes to actually getting started. Once we get to that blank page where we are supposed to write the prompt or start giving directions, it’s challenging to know where to start. As I previously mentioned, I find that it’s helpful to view an agent as an assistant, almost like a person who is hired to help you carry out various tasks. Your job, then, is to give that person clear and detailed directions on how to complete those tasks. An easy way to get started is with the “Create” mode. Instead of writing a prompt directly, you can approach it as more of a conversation. In your conversation, tell the agent its goal, give it a persona, instructions, etc, and it will automatically add it to the system prompt.
So now you’ve created something, however with a few tests you may have found that it doesn’t perform as well as you hoped. This is an issue that I often faced myself. When I finally completed writing out the prompt, the agent performed quite poorly. It almost made me think that maybe these agents really are not all that great or smart. I think this can discourage people who have higher hopes for what these agents can do. However, I discovered that with most agents, it usually takes many changes, tweaks, and iterations before you get something that is actually competent and useful. Similar to writing an academic paper or report, your first draft is never perfect, it takes a lot of editing, revising, to produce something that is actually good. The same principle applies when creating these agents.
Creating a useful agent is an iterative process. It takes testing, tweaking, rewording instructions, and occasionally starting over. That is not necessarily a failure. It is a natural part of the process, just as if you were training a new assistant that you hired. The goal is not necessarily perfection in version one. Many times, I’ve started building an agent and then completely started over with a fresh prompt, or reworked large parts of the prompt. Sometimes you begin with a very general prompt and update it as you test and use it, with gradual improvement over time.
Some points to consider when designing an AI agent
When creating your first prompt, either in a conversation with the agent, or by directly writing the prompt, consider each of these steps.
- Give your agent a persona. This is important because it gives the agent an identity. For example, “I am writing a book, you are my editor.”
- Give the agent a clear goal, or set of goals. For instance, “Your goal is to help ensure my book is well written by giving me constructive criticism.”
- Specify instructions or steps the agent should take. For example, “First, read through the book. Second, tell me the strengths and weaknesses. Third, flag any factual errors,” and so on.
- Provide examples of good and bad responses so the agent can model better outputs and avoid unhelpful ones.
- Give the agent context. This could be files, images, documents, or anything else the agent needs in its knowledge base. Think of this like uploading information into the agent’s brain.
- Specify the actions the agent should take, for instance, searching the web, writing code, generating images, and so on.
- Add guardrails. Since these agents are conversational, it is possible to pull them off topic. To reduce this, you can instruct your agent to avoid unrelated topics and stay focused on its work. This is not a perfect method, but it is a good practice, especially if you plan to share your agents with others.
- Test and iterate. Evaluate your agent, then revise it as needed. This is very important, remember, your agent will probably not be that good the first time.
A real example “The Quercus Companion”
I thought I’d share an example of an AI agent that I actually use regularly called the Quercus Companion. We at Rotman Digital designed this because part of our role is to help faculty and staff with Quercus related questions. Sometimes these questions take a lot of time and research, and testing to answer adequately. It isn’t always as straightforward as looking at the Canvas Guide to find the answer, especially for niche issues. Sometimes we find ourselves watching videos or reading blog posts to find answers. So we came up with this AI Agent to help us with the research aspect of some of these Quercus questions. Essentially, it functions as a research assistant.
Its goal is not simply to answer questions about Quercus. Instead, we designed it to first identify the problem, search for potential fixes or answers to the problem, use relevant documentation and support pages including U of T and Rotman specific resources where possible, posts from Instructure Canvas, blog posts from canvas, and many other sources from around the internet, summarize what it finds in plain language, and include links to the sources it used so I can verify the details myself.
In that sense, it does not replace the need to learn more about Quercus. It acts as an assistive research tool. A faster way to orient myself and track down useful information. And of course, I always double check the sources and confirm that the answers are accurate before acting on them.
What I like most about this example is that it is not especially advanced or fancy. It is fairly basic by design. The real difference comes from being clear and specific in the instructions, and refining them until the output is actually useful.
The big takeaway
If there is one thing to remember it is that you do not need to get it right the first time. AI agents are built through experimentation and refinement. The more you test and refine them, the more helpful they can become.
In future posts, we will dive into more practical prompting tips you can try right away. For now, if you are curious, start small, stay curious, and give yourself permission to iterate.
Please also feel free to email us at RotmanDigital@Rotman.utoronto.ca if you have any questions.