TMSAi Summer School 2025: AI Agents

If you've been following any AI news over the last year, you've probably heard about AI agents. In this post, I'll break down what agents are and how you can start using them to save time and enhance your workflows.

What are AI agents?

Few can agree on a firm definition of an AI agent. Here's my take: AI agents have agency. They can make decisions, use tools like APIs to connect to online services, and multiply the power of a single AI assistant in a coordinated team to execute tasks and deliver results at scale.

Let's explore a simple use case. If you send a large language model (LLM) an input like, "What's the weather today in New York?", the model will respond from its training data, but it won't actually have any context about today's weather.

However, if you give that LLM a tool to look up the weather - like the OpenWeatherMap API, AND tell the LLM about its available tools, then ask, "What's the weather today in New York?", it can decide to use its weather tool, get the API response containing today's weather report, and provide that information conversationally.

Today's reasoning models better understand when to use their available tools, but even smaller, non-reasoning models like Llama 3.1 can make good decisions about next steps in a task sequence—when prompted well and provided with a decent toolset.

So what? I already know how to check the weather.

Here's a more relevant marketing example.

Have you ever tried to get ChatGPT, Gemini or Claude to write an ebook, or an entire campaign's worth of content, in one shot?

It's pretty difficult to get the models to stay focused over long-running tasks. 

However, if you break up a task into steps, they excel in short term production.

Taking the ebook example: if you start with a planner agent to write an outline, introduce a search agent to retrieve info, spin up a bunch of writer agents to write each chapter, and add an editor agent to work on transitions and weave content together, you have a framework for writing long-form content. 

That's how I built an ebook writer app that features several AI agents. The finished product - around 4,000 to 5,000 words - is written and exported to Google Sheets within minutes, fully sourced and ready for fact-checking and editing. You can then add your perspective or insights from the subject matter experts you've interviewed. 

Human in the loop

AI agents are powerful, but they can't replace your critical thinking and judgement. 

As you start to create workflows with AI agents, consider the milestones where you should step in and make decisions.

When I built a small app that recast a TMSA YouTube video into a blog post draft, I initially had one AI agent analyze the transcript and choose a blog style - inverted pyramid, FAQ, listicle, etc. - before passing to another AI agent to write the draft.

When speaking with TMSA Executive Director Jennifer Karpus-Romain, she requested that we leave the style choice to the person running the app, rather than the AI agent. This way, content could be managed more strategically for the association.

I anticipate that we'll all face more of these types of decisions, as more AI agents enter our workflows. Recognize that - at least today - the more autonomy you give an agent, the less reliable that agentic flow becomes. That's because one error upstream can carry forward and cascade into additional errors downstream. So, be strategic about where you monitor the work to gain efficiencies while maintaining accuracy.

How to get started building AI agent workflows

Creating AI agents can still be pretty technical but the AI apps you already use are starting to introduce these features.

ChatGPT now includes their recently released GPT 5 model with various levels of "thinking" - more for complicated tasks, less for everyday tasks. ChatGPT is more like a model system now, rather than a bunch of individual models, since it dynamically chooses how much thinking it applies to your input. It also can automatically search the web to source information if it doesn't have the right context in its training data. The model makes choices and has access to tools - the two hallmarks of agentic behavior.

ChatGPT also features "Agent Mode," which combines two previously separate features: Deep Research and Operator (which starts up a virtual computer and takes action).

I find Agent Mode most useful for combining Deep Research with output creation, such as transforming the research into a report or a slideshow. The slides always need work, but the flow gets you further faster. Even shaving 25 percent off task time is meaningful time saved.

Google recently released Opal - an experimental app where you string together activities using their AI models. You can create a workflow with search, their text AI models, image generation, video generation and more, no code needed.

Let's quickly build an Opal workflow to create 10 versions of a LinkedIn ad, based upon text from a web page. We'll use copy from this TMSA page - the pitch to prospective Corporate Members.

1. Go to https://opal.withgoogle.com and sign in. It's free but requires a Google account.
2. Click "Create New".
3. Click "Add Assets" and choose "Text". We'll add some info about TMSA.
4. Next, click "User Input". This is where you - the user - will enter the web page URL to kick things off when we run the app.
5. The next step should be a "Generate" step, where we'll use the Gemini 2.5 Flash model to bring in relevant context from the prior two steps and create a market brief. The sidebar opens and here is where you'll enter your prompt. I built this prompt using the Instant Expert framework I covered in the last blog post.
6. The third step spins up a LinkedIn ad copywriter with another Generate step. The copywriter takes the "About" text, web page content and the market brief, then writes 10 LinkedIn ad versions.
7. Open the Theme editor, and enter a theme. I chose 'trucking' for this example. Opal will pick colors and might include an image.
8. In the sidebar, click Preview, then Start. Enter the web page URL for Corporate Membership, and we'll see the resulting 10 LinkedIn ads in a couple minutes.

Here are the results from my first run - pretty good for a first draft. Now we can enter other TMSA web pages, scaling up our campaign copy. 

Other no-code or low-code tools to create AI agents and workflows include Make, n8n,  Zapier and Lindy.

Using AI Agents for Research 

AI tools from the major model providers have revolutionized online research. What used to take hours or days now gets completed in minutes. 

Think of each of these platforms as an AI agent system. Here are the major AI research tools that I use, and when I use them.

Deep Research on OpenAI: Mentioned earlier, this is a nice research tool that can generate a thorough dossier in just a few minutes. Using this with a paid ChatGPT account can produce very helpful guides about your topic of choice.

DeepResearch on Google: With the Google Gemini app, I use their Deep Research feature for the deepest dives. Google’s competitive advantage lies in search, and I find that their tool, on average, delivers the best depth and scope of information across the web.

Research on Perplexity: I use Perplexity’s research tool for comparison shopping. They've invested in their ecommerce capability, and while not perfect, it tends to deliver solid results. 

Final Thoughts

As you may have heard about AI, it's as bad right now as it's ever going to be. In other words:  models, apps and ease of use will only improve from here. Salesforce has introduced agents, Hubspot has agents, and more use cases are being developed every day for sales and marketing.

Now that you have a foundation of knowledge about AI agents, what are you going to build? What else would you like to know about agents?

Connect with me on LinkedIn and let me know if you have any questions. In my next post for TMSAi Summer School, we'll cover AI and video.