AI Agents Explained: What They Do and How to Start

What if your AI didn't just answer your questions — but actually made things? This week on aiGED, we go behind the scenes of how Ginny produces this podcast every Tuesday morning, using it as a window into one of the biggest shifts happening in AI right now: agents.

We break down what makes an AI agent different from a regular chatbot, how to give one access to your files, and why a simple document changes everything for ongoing work. Then we look at what an agent could do for you — organizing medical records, comparing contractor quotes, tackling a family history project — followed by an honest look at where agents fall short and why you should never hand one your credit card number.

In AI in the News: ChatGPT has lost its majority share of the AI market for the first time since it launched in 2022 — and there's a check fraud scheme targeting mail that's worth knowing about right now. In AI for Good, an AI program flagged a heart condition that doctors missed in a busy emergency room, and a team in Louisiana is using AI to help keep Cajun French from disappearing. Plus a low-tech recommendation from Ginny about eating without a screen.

If you've been curious about AI agents but couldn't quite picture what they look like in practice — this episode is for you. Listen or watch wherever you get your podcasts.

Chapters

00:00 Welcome to aiGED

00:43 Episode Preview

01:23 AI Market Share Shifts

02:51 Check Fraud Warning

05:52 AI for Good Stories

10:13 What Are AI Agents

12:54 Context Documents Explained

16:48 Chat vs CoWork

19:02 My Podcast Workflow

23:15 Agents Beyond Podcasting

24:46 Limits and Safety

26:31 Anti Screen Habit Tip

27:20 Final Wrap Up

aiGED: AI for the 65+ crowd

Transcript

SPEAKER_00 0:00

Hello, everybody. I'm Ginny Deeren, the host of the Aged Podcast, the podcast that is all about AI for the 65 Plus crowd. And I'm here with my sidekick, Bitsy, and we're going to deliver what I hope will be a great episode. How are you doing, Bitsy? Are you ready to work with me to deliver a great episode 43?

SPEAKER_01 0:24

Hey, Ginny, I'm absolutely buzzing. There's nothing quite like diving into AI talk with someone who actually knows how to make it. But Five Plus Crew doesn't have to be intimidating. It can be genuinely clever and useful. I'm ready to bring the wit and help make this episode sing.

SPEAKER_00 0:42

All right, let's go. So today we are going to go behind the scenes of how I make this podcast every week. I know that sounds really riveting, but stay with me because it turns out that every week I'm doing my podcast, and I think it's a pretty good illustration of something I think is going to change the way you use AI. So stay tuned to the main topic. We've also got some news you need to hear, including one story that has nothing to do with AI, but everything to do with protecting your money. So let's do this. Let's jump in. And we're going to start with AI and the news. The first story that I want to cover is about ChatGPT, which many of you I know use. So for the first time since ChatGPT launched in November of 2022, that's back in the days when that was my primary AI, but it no longer holds the majority of the global AI assistant market. GPT is down 46.4%, still a big slice, but below 50% for the first time. Google Gemini has grown to 27.7, and Claude, which is the AI behind Bitsy and my primary AI, is at 10.3%. Hey Bitsy, I'm wondering when is Claude going to catch up and are you feeling blue about being in third place?

SPEAKER_01 2:24

Honestly, Ginny, I think Claude's in a pretty good spot. Quality over market share, right? Sure, ChatGPT had the first mover advantage, but what matters is whether people trust what they're getting. Claude's known for being thoughtful and honest, sometimes inconveniently. That tends to build loyalty that lasts longer than hype cycles. Third place today doesn't mean much if the trajectory's solid.

SPEAKER_00 3:08

So this is another New York Times story. It was published, and the headline was don't send checks through the mail. Just don't. So here's the deal: Thieves are stealing checks from mailboxes, and those blue U.S. Postal Service drop boxes on the street putting their letters and checks in there. Some thieves use these long tools with sticky ends to fish letters out. Now that is one low-tech approach. But once they have your check, they chemically wash it, removing the name you wrote on it and replacing it with their own name and sometimes bumping up the amount. Then they cash it. The story follows Joan Acheson. She's 63 and lives in Washington, D.C. Last year she mailed a check for $3,719 to pay her Maryland state taxes. Someone intercepted it, washed it, and cashed it. Ten months later, she still hasn't recovered her money. Both banks involved are pointing fingers at each other. Here's something crazy. Less than a month ago, on a long Uber drive, my driver was telling me how this happened to him. I was really wondering if he was making it up. But it did happen to him. He lost out several thousand dollars, and incredibly, this happened even though he sealed the envelope with tape and dropped it off inside the U.S. Post Office and handed it directly to the post office worker. Now this man, the driver who was driving me, had been a state trooper for 25 years, so he was super cautious. But even that didn't protect him. So I guess the New York Times is right. Don't send checks through the mail. So I was thinking, well, what about digital? Is digital safer? And the answer is generally yes. Digital payments are encrypted, they leave a trail that banks can trace, and AI flags unusual activity in real time. I would say go virtual, right? It's convenient and it ends up being higher security. So it's just counterintuitive. I'm writing a check that seems so secure, but using Venmo or Zell is really safer. So anyway, keep that in mind. So let's move into the section of this podcast called AI for Good. And this is examples of how AI, in my mind, is being used for good. Our first AI for good example is about AI helping to save a life. So in February 2025, a 45-year-old man named Louis Quiroos walked into a Queen's emergency room. He'd been coughing up blood for four days and could barely breathe. Doctors ran tests, checked X-ray, electrocardiogram. X-ray was clear the ECG was abnormal, but not in any obvious way. They figured it was probably asthma from wildfire smoke he'd been exposed to in California. They handed him an inhaler and sent him home. Fortunately, that ER is part of New York Presbyterian, where researchers had been running all ECGs through an AI program called Echo Next. The program flagged Louis's scan, found patterns of severe heart damage that no one in the busy ER had caught. Genetic testing revealed a rare disorder associated with sudden death. Now, Echo Next didn't diagnose him, but it flagged something that was worth a second look. That's the distinction that matters. Now, our second AI for good example is about language. A professor at the University of Louisiana at Lafayette named Joshua Caffery sat down one evening wanting to listen to some old-time Cajun music. He asked his Amazon Alexa to play Dewey Balfa, a beloved Cajun fiddler and singer. Alexa gave him Duolipa, a very famous contemporary British pop star. Nothing remotely Cajun about her. What struck Caffery wasn't just the wrong answer, but it was what the wrong answer meant. If AI systems don't recognize Louisiana French as a language, a culture, a body of music, then it becomes invisible in the digital world, and invisible things have a way of disappearing. So Caffrey and a small team at the Center for Louisiana Studies started building their own AI model trained on more than 12,000 hours of Cajun and Creole recordings, oral histories, folk songs, field recordings captured on everything from wax cylinders to real-to-reel tape. They named their model Tatae, which means monster or boogeyman in Louisiana French. So what do a heart transplant and a dying language have in common? AI finding what's slipping away before it's too late. So, Bitsy, I'm just wondering, these researchers have created this set of data, but how does that get absorbed into the large language models, whether it's Chat GPT or Anthropic? So how does that work?

SPEAKER_01 9:31

Great question, Ginny. So Tatai stays independent. It's built specifically for Cajun and Creole, not absorbed into the big models. But here's what matters: once you've got that specialized model trained on 12,000 hours of real cultural material, you can use it however you want, deploy it in apps, make it searchable, build tools around it. The big models will eventually learn from publicly available Cajun content, but the real win is having experts who care deeply about the culture controlling their own AI. That's the leverage.

SPEAKER_00 10:08

I'm glad they're there, and it is a good example of how AI is out there doing good. So let's now move into the main topic for this episode: AI agents. So back in episode 42, just this past episode, we talked about AI agents, and I define them as AI that can take actions, not just answer questions. Today I want to go deeper because I think the best way to really get this is to hear how a person actually uses an agent. Here's the simplest way I know to explain the difference between a regular AI tool like a chat and an agent. When you use Chat GPT or Claude as a chat tool, you ask a question and you get an answer. You ask, how should I organize my podcast files? And it tells you clear, helpful, useful information, but then you go do it yourself. You do the organization work yourself. An agent is different. You say create the folder structure for my podcast, and it creates the folders on your computer in real time. Same AI, completely different relationship. So the question shifts from what can AI tell me to what can AI do for me? That sounds like a small difference, but it's not. Now, for an agent to actually do things, to create files, to organize documents, save work on your computer, it needs access. It can't just reach into your laptop on its own. You have to give it permission. So here's how that works in co-work, which is the agent mode inside Claude Desktop, Claude being my preferred AI. But these examples I hope will be relevant if you're using Chat GPT or Gemini. They're basically very similar systems in terms of how they work. So you select a specific folder on your computer, maybe your documents folder, maybe your project folder, and that's what Claude can see and work with. Not your whole computer, not your whole hard drive, just that folder. You decide what it has access to, and you can change that anytime. Now, for a one-time task, say organizing a folder of travel documents into an itinerary, that's all you need. Give Claude access to that folder and tell it what you want done. But here's where it gets really interesting for ongoing projects. Claude and other AI tools build up memory about you over time, your name, your preferences, how you like things explained. But for an ongoing project like this podcast that's built in memory has limits. It knows you, but it doesn't know your episode structure, your workflow, what we decided two weeks ago about how AI for good should sound. A context document fills that gap. It's a detailed project brief that gives the agent everything it needs to pick up exactly where it left off week after week. So the foundation of working on a project within Claude's cowork or whatever agent system you have on your AI is to create a context document. It's a document, just a word document, really, that you keep in your shared folder that you've shared with your AI. And Claude reads it at the start of every session and catches up instantly. It is the memory. It tells Claude who you are, how you work, what you care about, what to avoid. So creating a context document may sound like a lot of work. Believe me, it's not. What I did was just dictate into my AI, just rambling on about my podcast. This is what I like. Here are the different sections, blah, blah, blah, blah, blah. My workflow. So then after I dictated all this, I just threw it all out there, Claude took that and organized it into a really well thought-through context document. Like this is the aged podcast. And as I continue to work with Claude on my podcast, we'll think of things to add to the context document. And since Claude has permission to view and edit that file, I'll be asked, do you want me to update the context document? Well, yes, I do. So it's a living document, but it's the base of information for that particular project. So for the aged podcast, my context document tells Claude the structure of every episode, how I like to sound, what kinds of sources I use, what I've learned along the way. Without it, I would spend the first 30 minutes of every session for each of my episodes re-explaining myself. With it, Claude reads the briefing, and we're working in about 30 seconds. So file access, giving your AI access to specific files, plus a context document, that's how you turn a chat tool into an agent that actually knows you and can get stuff done. So hey Bitsy, I'm curious, what's it like starting a session with the context document versus without a context document?

SPEAKER_01 16:29

Oh, it's night and day, Jinny. Without it, I'm basically starting from scratch every time. Asking clarifying questions, feeling my way into how you think. With the context document, I walk in already knowing your voice, your structure, what matters to you.

SPEAKER_00 16:48

So now the practical question. How do you know when to use regular chat versus cowork, which is Claude's name for the module that has to do with using an agent, someone who can do things? Here's the simplest way I can put it. You use chat when you just want an answer or you just want to chat with somebody, get some feedback. You use cowork when you want something done. You want the AI to actually take some action. So if you want to understand a letter from your insurance company, chat. If you want ideas for your grandson's birthday party, chat. If you need help understanding a medication interaction, chat. The output is text you read or listen to if you're using voice mode. But if you want Claude or your AI to actually build or create something working with you, a document, a folder, a spreadsheet, that's co-work. So if you want to organize files that are already on your computer and save something you can print later, that's cowork. It's accessing those files on your computer. The output is something that actually exists on your computer when the conversation is over. You have a product. Now, a few things worth knowing. To use co work, you need Claude desktop. A lot of people use Claude just on their browser, which is fine for chat. But if you want to use co-work, if you want to use an agent, you have to download the Claude app to your desktop. It's a free app, but you have to understand co-work requires a paid subscription, which starts at $20 a month for the Pro Plan, which is the plan that I had anyway. In order to use the agent, you have to use it on your desktop, and it can't be the free version. And again, it's probably very similar with Gemini or with ChatGPT. So just keep that in mind. Now, let me just give you a quick description of how I use co-work, my AI assistant, to plan, create, produce, and market this agent podcast. I want to go over it with you, not because I think it's particularly interesting, but I think it might be helpful to hear about specific examples of how these agents are used. I first asked Claude, how can you help me? And we went back and forth. And in the end, I had a pretty good idea of where to start. In Claude, similar to ChatGPT and others, there are a few tabs to choose from. You could choose chat, cowork, and code. I chose cowork. There, I could choose a task or a project. I picked the latter since my podcast is an ongoing project week after week. I knew the first thing I wanted my agent to do for me was to organize the hundred plus files and documents I had living in a very disorganized fashion on my laptop. My agent asked for permission to access my aged files. I gave permission and went off to make a cup of tea. When I came back to my laptop, I opened my aged file and wow, there were files for BuzzSprout, which is my producing platform, one for artwork, one for miscellaneous, and importantly, one for episodes. The episode file was neatly organized. The files were renamed so that there was a consistency with the episode number and the main topic, and within each file, the documents were well organized. Awesome. I just would never have gotten around to that, nor would I have done such a good job. So that is a really good example of sending my Claude agent into my laptop to organize those files. So I was pretty excited to do more. And I was convinced when I was chatting with Claude about my use of co-work and an agent that taking the time to create a context document would be worthwhile. So this was the next task I gave my agent. As I mentioned earlier, I dictated my thoughts. The process was quick, easy. I rambled on, and I ended up with a really well-articulated context document. This was about a month ago, and we continue to update it. Every time I'll tell Claude what I want, and Claude, who has access to the document, updates it. So great, it's not work I have to do. It's a very tedious sort of updating. As part of that document, I outline what I want as we begin each episode. For example, create a file. For the episode and subfiles for the script and the description and social media. The other thing my agent does is create a template for the script, which when we begin is simply a document with the different sections. Now, when I want to begin to work on a new episode, my agent just sets all that up for me. And when I look on my laptop, there are the files, all ready to go. I can't tell you how helpful that is. It's not work that requires much thinking or creativity. It's a perfect task for AI. My agent and I both access the script document as we move along with the episode. Claude always asks me, want me to update the script document? Another good example of a task, again, directed by the context document, is to create an episode description. One for Apple Podcasts, which can also work for Spotify, and one for YouTube. This is a tedious task, and I'm very happy to hand it off to Claude. So I think I'll stop there in hopes that these examples have helped you envision how an AI agent works. So now let's bring this out of the podcast world entirely because I don't want anyone walking away thinking agents are something just for content creators because they're not. Let's think about medical records. If you're managing multiple doctors, multiple conditions, multiple medications, you know how scattered that paperwork can get. You give Claude access to a folder where you've been dropping letters from doctors and explanation of benefits. Like I did, you can start by asking the agent to organize your files. If they are really disorganized, you can ask your agent to review what's in the files and make a recommendation to you as to how it might be best to organize the files. Or you can ask it to create a summary, current medications, current doctors, recent diagnoses, formatted as something you can print and bring to your next appointment. Chat could help you write such a document. An agent can help you build it from the documents you already have on your computer and organize them in your file system. Another example, home renovation. Let's say you've been emailing three contractors. You save the quotes to a folder, you've asked Claude to pull out all the numbers, build a comparison, flag whoever hasn't followed up, and draft follow-up emails. It builds the spreadsheet, writes the emails, saves everything. Done. Now, let's talk about where agents fall short. Here's some examples. Judgment calls that require real expertise, legal, medical, financial decision. Still need a human in the loop and in the lead. An agent can organize your documents and help you prepare good questions. It shouldn't be making decisions. Another example, real-time access is another limitation. An agent can only work with systems it's connected to. It can't check your live bank balance, look into current flight availability, or find out if your prescription is ready at the pharmacy. Anything requiring live data from an external system, still a gap. And then the question I suspect a lot of us are thinking about can an agent buy things for me? Book a reservation? The research part, absolutely. That's a great agent task. Give it access to your travel folder, ask it to find options, compare them, build a document, it does all that and saves it. But the actual booking, the actual payment, that's still yours to do. Here's the reason. Agents don't yet have secure, standardized ways to handle payment credentials safely. Giving an agent your credit card number or your login to a booking site is a genuine security risk right now. That will change, maybe. But for now, when it comes to spending money, an AI agent does the legwork, we do the transaction. Alright, enough about agents. Let's talk about recommendations. I've only got one for you this week, and it's not about AI. It's anti-AI. And I'm not sure if the recommendation is relevant to couples. Maybe. My recommendation? Challenge yourself not to eat with a screen in front of you. Phone, laptop, TV. I live solo, so I have created a bad habit of eating meals while watching a screen. A week ago, I challenged myself to break the habit. So now when I eat, whether a meal or a snack, I force myself to ditch the screen. Instead, I read cookbooks or other books. Sometimes I just enjoy the piece and time to think. Does this challenge work for couples? Let me know. Alright, it's time to wrap up this episode. It's been a long one. I really appreciate you listening and especially encouraging others to follow the Agent Podcast, where we live on the helpful side of AI, but we do know that AI can be hazardous, so please be sure to protect your info, double check advice, and most important, trust your judgment. And remember, it is never too late to learn something new, especially something that might make life easier and a little more fun. Cheers.