Ep 849: AI Agents in 2026 Explained: What They Are and When You Should Use Them (Start Here Series Vol 8)

Welcome to the Everyday AI Podcast. My name is Jordan Wilson, and for the past 3 and a half years, we've put out more than 800 episodes. Yet one of the most common questions I get, I didn't really have an answer for. Where do I start on the Everyday AI Podcast? And that's why we started the Start Here series. And with fall now back in full swing, the Everyday AI Podcast is going back to school and playing back the entire Start Here series. From front to back. We've hit pause on our normal Monday to Friday programming to run back our most popular series ever for the next 30 days.

We made the Start Here series for beginners and AI champions alike.

So whether you're just trying to get a grasp on large language models or grappling with the best coding harness for multi-agentic workflows, the Start Here series covers it all. Plain language, no jargon, and easy to follow along each day. So make sure to subscribe to the podcast and check back each day for new insights, day by day. The series is a culmination of spending more than 10,000 hours covering generative AI over the past 3 and a half years. So you don't want to miss a single episode of the Start Here series. Let's get into it.

Whether you're newer to AI, or if you consider yourself an AI expert, we're all probably thinking the same thing. AI agents, AI agents, AI agents. That's because, well, we've been promised that AI agents would take over, maybe in a good way, maybe in a bad way. Yet that promise never really came to fruition. I mean, we heard it in the early ChatGPT days, 2023, 2024, 2025 was supposed to be the year of the agents, but it wasn't. But guess what?

We're there.

I think we've actually arrived with some of the recent AI updates from the big companies in Anthropic, OpenAI, Google, Microsoft, even crazy viral open source projects like OpenCLAW. I think now the long vision that we've been pitched of AI agents is here, but it's all happening at once. Like I said, I've been doing this everyday AI thing for 3 years, and at least when it comes to AI agent development, the last month has been as eventful as the other 2.9 years combined. Don't worry, even if you are brand new or if you're an expert. That's what we're going to tackle on today's show. We're going to hopefully get you caught up, tackle AI agents from all angles, and do so in a Hopefully quick, beginner-friendly way. So this is the Start Here series, and we're going to be going over AI agents in 2026, what they are, and when you should use them. All right, I'm excited for this one. I hope you are too. What's going on? Welcome to the Start Here series. My name is Jordan Wilson, and the Start Here series, well, I created it because after 700+ episodes of the Everyday AI podcast, I didn't have a good answer for the most common question I got. People say, hey Jordan, looks like a lot of great information here on your podcast. Where do I start? Well, you start here with the Start Here series, and this is volume 8 in the Start Here series. But this is an essential podcast series to both learn the AI basics and to double down on your AI knowledge. And the way that you should really do that is make sure you go to starthearseries.com. That is going to give you free access to our inner circle community in our Start Here series space. So you can go catch up on all of everything in this series in one place, a bunch of additional resources, resources that we've thrown together, all available for you right there. So like I said, whether you're brand new, an expert, it doesn't matter. Go join now like 1,000 people that are already in the community and listen to all of the Start Here series. Also, I have to put this out there because it's important. Make sure you also go listen to episodes 712 and 713. That is our 2026 AI predictions and roadmap series. You know, 26 kind of bold predictions on AI, and at least a third of them are about agents.

All right.

So, and if you missed our last Start Here series, that was volume 7, where we went over context engineering, how to get expert-level outputs from AI chatbots. But let's get straight into AI agents, kind of the state of AI agents here in 2026. So here's what we're gonna be going over on today's show. We're gonna talk about what AI agents actually are and how they're different from traditional AI chatbots or AI-powered automation. I'm gonna give you a decision framework for when agents are the answer and when they're worth it versus when they're just overkill and when you should say, hey, that should just be a human task or an AI, you know, large language model task. And then I'm gonna give you the practical path to safely starting your first agent pilot, not even this year, like this week. All right, let's get the explanations out of the way. All right, so like I said, if you are— if you've been building AI agents for, you know, a decade or 2, you can just skip ahead a couple of minutes. Actually, speaking of that and why I think it's so important to look at this from a foundational level, so about 2 weeks ago, had a fantastic guest on my podcast. She's the head of Microsoft Research. She's been working in AI agents for 20 years. And one thing we talked about both on the show and before and after is the definition of an AI agent is constantly changing, right? So even what is an agent and what's not an agent, what can an agent do, what can't it do, it's always changing. All right. So if you're listening to this in, you know, February of 2026, That's great. All this information is probably going to be extremely fresh and accurate. If you're listening to this in June or in 2027, right, some of this might be a little old. So keep that in mind. But I'm going to give you at least as of February, you know, 19th, 2026. This is the realest, most up to date. So what is an AI agent in very simple terms? It's kind of like an AI chatbot, but it has tools and it has permissions and it has a goal. And it makes its own decisions. And sometimes it might start off on one path and say, this isn't right, and it'll go backwards and start down another path. So they can plan ahead, take steps, use tools, take actions, check results, and retry within the guardrails, whatever those guardrails are, whether you set them up or they're set up by a certain, you know, piece of software that you're using. So what do agents do? Well, they can take on delegate work. They can delegate work to other subagents. They don't just answer questions, right? So they can draft, route, research, navigate apps, and even compile multi-step tasks across different systems. And here's the state of AI agents in the enterprise right now. Gartner studies show that 80% of Fortune 500s have active agents right now. Sorry, that was from Microsoft. But Gartner said that 40% of enterprise apps are projected to include them by 2026. So yeah, even since that story— or sorry, that report came out from Gartner last summer. I think that has changed. That's one of the things that's difficult to give people a thumbprint on the state of AI agents is because even the studies that are the most up to date are usually looking in the rearview mirror a couple of quarters or more. But the reality is an overwhelming majority of Fortune 500 companies have agents in production. All the software that you use anyways, because you might think, oh, an agent means building something on your own. No, it doesn't. Right. And it's not just the big 4 in Microsoft, Google, OpenAI, and Anthropic, right, that would give you kind of agents out there in the workforce. There's literally thousands of different now pieces of software. And that's where this, you know, Gartner stat comes from, that 40% of enterprise apps are projected to include AI agents this year.

Right.

What that means, a good example I've talked about, I've used, you know, ClickUp, a project management tool in the past. Right. So now like ClickUp, has agents by default. I think things that used to be considered just features or tools in daily software are now agents. If you're using Salesforce, they have agents. If you're using Slack, Slackbot, they have agents. If you're using HubSpot, they have agents. It's not just building an agent from scratch or using one of the tools from the Big 4 or these super general agents like Meta's Manus agent. That's weird to say now, right? Manus that was acquired by Meta or GenSpark or some of these, you know, general agents, right? They're everywhere. So what is it actually, right? So I'm going to break it down in the most simplest, hopefully, language possible. An AI agent is an autonomous system that can plan, execute, and self-correct across different tools and systems. So it's kind of like a human at a computer.

Right.

So that can make decisions and has access to things. That's the way I like to think of an agent. Right. It's the same thing as if a, you know, an entry-level worker sat down at a computer at your company that had access to everything. Right. So think of that. But you can have a million of them. So what feelings do you get? Is that excitement or are you thinking of security and risk and things that can go wrong? Well, I think when you think about agents and their implications, you have to think of both sides of that coin, but that's essentially what an agent is, right? A full agent is like sitting a human down at a computer that has access to everything, right? It can go get you new clients and close deals. It can complete work autonomously— spreadsheets, PowerPoints, uh, right, jumping between different apps. It, you know, has access to a terminal, uh, you know, a sandbox, a virtual browser, all these different things. Like a human sitting down in front of a computer does. So essentially, if it can't take action, if it can't adapt to context, if it can't solve new problems independently, it's not technically an agent. And I'm gonna go over some of the, you know, other things that it might be. And that's because I've talked about this on the show before. A great, a great study from Gartner talked about this concept of agent washing. Right. So they looked at thousands of different vendors in 2025 that were essentially marketing or offering or selling agentic capabilities. And what they found, only 130 of thousands were actually agents. It's agent washing, right? Most are just using, you know, chatbots or, you know, scripted, rigid workflows and sprinkling some you know, quote unquote, agentic model somewhere in the process and saying it's an agent. I tell you this, even looking at my own inbox, right, for Everyday AI, because we get pitched more than 1,000 times a year by companies. And, you know, obviously recently it's just been all agents, right? Oh, check out our agent. It's the best one in the world. Right. And most of the time I don't look because it's— I don't think most of the companies are going to make it very long. But then when I do look, I'm like, this isn't an agent. This is like a workflow. This is RPA with a little bit of a large language model sprinkled on top.

That's not an agent.

An agent, you don't give it a dedicated workflow. You don't give it if-then or if-else statements to follow. An agent is sitting a computer, like sitting a human down in front of a computer. They have free will, make decisions, make good ones, make bad ones. It really depends on, well, the guardrails that you set up around those agents. And right now, 88% of executives are investing in agentic AI, but most can't even tell what's real from fake. So let me quickly talk about the difference, right? Because I said chances are, even if you're paying for an AI agent product, it's just marketing. It's smoke and mirrors. Because like I said, according to that Gartner study, less than 5% are actually AI agents. So let's first talk about AI-powered workflows, right? And there's nothing wrong with these other things, FYI, right? But a great example of that is like N8N, right? Or make.com, or, you know, in some instances like Zapier as well, right? I love— I love these tools. I love these platforms, right? I'm not saying that they're bad, right? But if you have all these decision trees and you're connecting these nodes, it's not an agent, right? It's an AI-powered workflow. Nothing wrong with it. Alright. We're just getting definitions here.

Alright.

Then you have agentic models. Well, what the heck is that? Well, these agentic models are essentially turning into mini agents, and they're technically, I think, in my opinion, more agentic than AI-powered workflows because they can make more free will decisions on their own and also in the same time do more harm. Right? But agentic model. Well, that's just our You know, frontier AI models, right? Think of the big four models. Well, technically the big three that are at least creating and you know putting their models out there. You could throw in you know XAI's Grok if you want to. I wouldn't necessarily. But let's look at three models: Google Gemini 3 Pro, Claude Opus 4.6, and OpenAI's GPT 5.2 Pro.

Right?

Or five-three codex, whatever you want. And obviously, if you're listening to this, you know, at some point later in the year, it you know just replace it with you know today's three models. These models by default are agentic, and they didn't used to be. Right? A year-ish ago, models were not agentic. They were extremely smart autocomplete token munchers. Right? Round round round round tokens here spit it out. Right now they're not like that. Models are agentic by default. That means you can give them a lot of context. You can give them files, you can give them problems. You can even in the course of a chatbot, a front-end chatbot, you can technically, if you know what you're doing and give it enough context and go back and forth with it, it can technically have a very agent-esque workflow. It's a little more confined because you're working within the constraints of a certain AI chatbot like ChatGPT, but I'm talking front-end. Front-end AI models technically are very agentic because they can decide on their own how they're going to solve a problem. They have tools that you probably don't even know that they have that they will call on their own. They have virtual browsers. They have their own sandboxes where they can write and run code. Crazy thing I was doing recently using Codex. Very good. But I mean, technically, I think OpenAI is investing more and more into Codex. You know, it's technically a coding model, but it's a front-end model, right? You don't have to do anything on the back end. I was using it, and Codex on its own decided it was going to download a 6 gigabyte open source model, and it used it in its own platform. That is crazy, right? It noticed that it didn't have the capabilities to do something at scale. That I asked it to, and then over the course overnight, it worked for almost ten hours. It downloaded an open source model, and then essentially I had this one was had like three thousand screenshots, and it wasn't able to you know get the results it needed to at scale at speed with its own computer vision model. So it said, all right, I'm going to download something, and a six gigabyte model, and it ran it through there. Right? That's an agentic model. It made decisions. It called tools. It found solutions. Right? Then you have passive agents. All right. Sorry. I went on a little side tangent there. So number 1, AI-powered workflows. Number 2, agentic models. And then number 3, passive agents. So these are agents that are not necessarily autonomous. Right? But same thing. An example of that.

Right?

ChatGPT has a great agent mode. I can schedule that agent to run.

At 9 a.m., right?

But I still have to trigger it. I still have to go say, "Hey agent," right? And I can train it and give it you know its its role and its its context and all this stuff. But I still have to go push the button. All right. And then you have autonomous agents or always-on agents, right? And then these are taking especially recently a lot of new shape and form. But I mean, an always-on agent or an autonomous agent is essentially that it's working. Around the clock, right? And it's not just triggered by you. It can, in theory, be triggered by a certain action like, oh, when you receive an email, right, it's going to automatically do A, B, and C and then go after a certain goal, right? So even the lines, you know, kind of going over these 4 different classifications, they're starting to blur as well. And the risks, just like the capabilities, Are growing just as fast, right? Some some good stats here. So this is from Cybersecurity Insiders said that 75% of CISOs have discovered unsanctioned AI tools already running in their environments, right? So that's that's crazy, right? When we talk about shadow shadow AI, it's no longer just oh someone's using you know ChatGPT maybe when they shouldn't. No. What about if someone's building agents that have access to all of your data? It's happening. And then those agents are making decisions potentially without any or, you know, company-approved guardrails. And then when you talk about multi-agent systems, that's when there's huge new risks, right, that you— it's, it's really hard to project because as agents' capabilities are changing, Humans necessarily, even the humans building them, don't fully understand the risk profile of these multi-agent systems. A very simple example. I didn't know that Codex was going to download an open-source model. What happens when we set up guardrails on maybe a certain multi-agent system and then the agent gets creative? It finds loopholes or gray area, right? And it says, hey, well, it said I can't set up, you know, my guardrails say that I can't use other agents, but it doesn't say I can't use, you know, AI-powered workflows. So then it finds a loophole and all of a sudden you have these agents that are setting up something that is, you know, unregulated. Agents will do that. Agents by default are made to be, quote unquote, helpful assistants, just like large language models. So if they think that it's helpful, They're not going to, you know, go through a morality check first or a company ethics check unless you really have that hard baked in there. They're going to go off and make trouble, probably. They're going to break things while getting something done. You know, sometimes I think of agents, it's like giving a toddler a task, right? Hey, child. Hey, 3-year-old. You know, go, go put this toy away. All right. Well, did they put the toy away? Sure. Did they knock over a lamp? Maybe. Did they smear marker on the new couch? Potentially. Right. It's the same thing, you know. And right now, another report, the 2026 CISO AI Risk Report, said that 92% of organizations right now lack full visibility into their AI identities. And again, as those AI identities and capabilities are growing, the risk is going crazy. So how did we get here? Right. Let's take a very quick walk. So 2022, you know, AI was, you know, hey, type a question, get an answer. And then I think we had 2 years, you know, in 2023 and 2024 where, you know, large language models, oh, they can use tools now and you can upload files and it can, you know, browse the web and that's cool and it's more helpful. But then I think toward the end of 2025, and obviously this year, now this is where these reasoning models are agentic by default and their capabilities are growing by the day. And like I said, the flip side, the unknown risks are also growing. But now reasoning models and reinforcement learning just let agents actually act, right? And even if we talk about things like OpenCLAW, right, which OpenAI acquired, You know, these open source agents, people are just giving them computers. They're handing over, you know, their complete lives, right? Hopefully, maybe those are more entrepreneurs, business owners, solopreneurs, right? But people are still doing this. They're giving AI agents access to everything, and we don't yet know their capabilities. But it's quickly gone from AI is just a chatbot that you talk with to, oh, now all of a sudden that chatbot has access to tools and the internet. And oh, now all of a sudden, you know, these agents are making decisions completely on their own and their capabilities are quick. So here's 6 different agent types that I think you'll encounter in 2026. You know, not every single agent out there, like I said, there's literally tens of thousands technically agents because you can spin one up in a couple of minutes.

Right.

But you have task agents. These are agents that draft, summarize, create assets, etc. You have decision-supported agents. These are ones that compare options, surface trade-offs, and flag risks for review. Then you have process agents. These are ones that route work across tools, gather inputs, and prep the next step. You have computer use agents. These can navigate websites and apps to complete multi-step tasks. You have multi-agent systems, right, when agents can spin up sub-agents, you know, send a bunch of agents out on their own, right? Claude Code, you know, from Anthropic recently with their Opus 4.6 kind of, you know, have popularized that on the consumer and the mainstream. Then you also have commerce agents, right, where agents can talk to each other. Right. You know, agents are now having their own version of the web. Right. There's going to be agentic transactions that really don't involve humans at all. So even the type of agents changing quickly. Right. Talk to me next week. There's probably going to be new categories. So when do you actually need one? When do you need an agent versus when should I go do this task myself versus when do I need an AI-powered workflow versus when should I just use ChatGPT or Gemini or Claude or Copilot versus a full autonomous agent? Well, I think if you can write a clear checklist and it rarely changes, a workflow can handle it fine, right? AI-powered workflow. If the steps vary, if it requires judgment, or if it depends on multiple tools, That's probably when you're starting to get into either agent territory or an agentic model that has access to your data. But here's the thing. If the underlying process is broken before the AI touches it, an agent is not going to be the answer. I think a lot of people— well, by a lot of people, I say the overwhelming majority of people are just looking at an AI agent as a shortcut to do things that humans couldn't or to do things that they just don't want to do. It's not going to work like that. If you stick an agent on a broken or an antiquated workflow, you're just asking for a compounding disaster, right? If you're like, oh, well, our humans aren't doing a good job at this, let's give it to an agent. Well, the agent is going to do a worse job than the humans that weren't doing a good job. So you should really only be applying an AI agent to processes that are already running very smoothly, that are well documented, right? You have SOPs, you have ways to measure them. Otherwise, setting an agent out is just a recipe for disaster. That's why another recent Gartner study said that 40% of agentic AI projects are expected to fail or be canceled by 2027. Well, that's one of the big reasons is what I just told you. People are thinking of AI agents as duct tapes instead of being multipliers through processes that are already AI-enabled and working with humans. Poor data quality, just like anything else. Missing context, no evaluation loops, because I think success requires rethinking and redesigning processes, not just adding AI on top. You're probably tired of me saying, you know, you don't just upskill, reskill, right? The same thing with agents. You don't just upskill a workflow with an agent. You don't just upgrade a workflow with an agent. You have to deconstruct it, rebuild. You have to unlearn and you build from scratch an agentic-first workflow. So we'll talk about bounded autonomy. And that's, well, it's how you start losing control if you're not doing this. So you should be starting when it comes to agents. Like I said, pick the right workflow that's already working AI-enabled with humans. It's documented, it's processed, it's processing correctly. You can measure it, etc. So you should start at suggest only with an agent where the agent just drafts and you decide what happens next. Once it's gone through that process successfully, then you move on to execute with approval where the agent acts after you one-click sign-off. Then again, then, and this is a big asterisk, right? Everyone's got to sign off. That's when you can start moving on to more autonomy, right? When you can scale if the guardrails are right, you know, with spending caps, permission rules, and audit trails. All right. But only then, right? Can you think of AI agents as this, you know, infinitely, infinitely scalable asset for your company? So here's how you get started this week. Number 1, you need to think about meetings to action items. Have an agent go draft owners and follow-ups and you approve before sending. Inbox triage, right? Another great thing for AI agents. Don't have them act, right? Have it triage your inbox or multiple pieces of software that you already use. Or creating research briefs. An agent can build a briefing doc with sources and let you verify them before sharing. So start Very simple. Don't let an agent do something that right now requires multiple humans and multiple approval processes. Start looking at maybe some of those time-consuming, difficult tasks that an individual human does, and then don't give the AI agent full autonomy. Go through steps. So here is the big takeaway as we wrap up. Agents are for delegation, not just answers. They plan, Act and report back. They can use tools at their own disposal, come up with new solutions that you didn't even think of. So that's why guardrails, traceability, and observability are paramount. Then you need to start with bounded autonomy and measure time saved, error rates, risk events, you know, ROI. You need to constantly doing those things. You need to don't do human in the loop. Right. It's one of the biggest mistakes I think companies are still making because human in the loop was this fun term in 2024 and everyone just said it. No, human in the loop is combustible in a bad way. Right. People think, oh, we're going to have 100 agents and we're going to have, you know, Bill and IT oversee them. No, you need expert-driven loops with cyclical improvement chains that are chains that are documented.

Right.

Observability, traceability is a full-time job for a team. Go back and that's why I said go back and listen to the 2020 6 AI Prediction and Roadmap series, I talked about this in a lot more depth and told you how to get it done. And then last but not least, the future advantage comes from building agent ecosystems, not just picking the right agent or, you know, dipping your toe in the AI agent of the week. You have to be intentional about it. And investments now, even sandboxing them, not in production, investments now in proper AI agent usage are going to be key for when the technology improves. I know it's a tired old cliché, but AI agents are the worst— are worse today than they'll ever be. They're only going to get better. All right. I hope this was helpful. A beginner's look at agents. So whether you are brand new to AI and you're getting all excited about all this, all these shiny capabilities, or if you've been building AI agents for a decade, I hope that this episode and also the Start Here series has been helpful. And if it is, please, please, please go to starthearseries.com. That is going to give you insider access and free access to our inner circle community. And it's going to spit you right out once you join in the Start Here series space there. So you can go watch, listen, read more about all of the Start Here series in order. Connect with other people who are doing the same. All right. I hope this was helpful. Thank you for tuning in. Hope to see you back tomorrow and every day for more Everyday AI. Thanks, y'all.

And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit youreverydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers, and we'll see you next time!