2025 was the year of agents, what's coming in 2026?

0:03Welcome to the Practical AI Podcast, where we break down the real world applications of artificial intelligence and how it's shaping the way we live and work. and create. Our goal is to help make AI technology practical, productive, and accessible to everyone. Whether you're a developer, business leader, or just curious about the tech behind the buzz, you're in the right place. Be sure to connect with us on LinkedIn, X, or Blue Sky to stay up to date with episode drops, behind-the-scenes content, and AI insights. You can learn more at practicalai.fm. Now, on to the show.

0:48Welcome to a new year of practical AI and an episode with just Chris and I where we try to keep you fully connected with everything that's happening in the AI world, which is a lot these days, both last year and this year. But I'm Daniel Whitenack. I am CEO at Prediction Guard, and I'm joined as always by my co-host, Chris Benson, who is a principal AI research engineer at Lockheed Martin. Happy New Year, Chris. Hey, Happy New Year, Daniel. It's 2026, probably the fastest moving AI year ever coming up here. First, well, every new year, I guess, has been the fastest movie. Well, I guess since we started the podcast, you know, whatever, eight years ago.

1:39It was a safe thing for me to say. There you go. Yeah, yeah, yeah. Safe thing for you to say. I mean, granted, these last few years have been a little bit frantic in relation to the years prior to that with the podcast, which felt, you know, in retrospect, seem a little bit chill. but it definitely seems like 2025 was a big year. 2026 will be a big year. And so as we're coming into the new year for our listeners, usually we try to do some type of, we don't have a strict format here because we're pretty casual, but some type of discussion of things that happened in 2026, themes, looking for things that happen in 2025.

2:30I'm already a year ahead, I guess. Things that happen in 2025 and things that may or may not happen in 2026. Usually our predictions are wrong, as are all predictions. I'm okay with that. All models are wrong, but hopefully this podcast will be useful. Yeah. So interesting times, Chris, interesting dynamics in our world in all sorts of ways. But if we hone in on AI, I think at the beginning of last year, if I'm remembering correctly, there were a couple of things we talked about. One of those things, or certainly at least a theme that we've talked about a lot this year, which if we were to categorize the year 2025, I don't know if you would agree, Chris, but it does seem like the year that we transitioned to talk about AI agents.

3:21It was sort of like for a while we talked about models. Then we kind of talked about assistants. And then we really kind of transitioned to talking about agents. Agents are autonomous AI. That was a key theme of 2025. I guess one first question, Chris, what did we actually do with AI agents in 2025? Overall, was it a positive and or successful year of trying agents? Well, I think there was untold levels of hype around agents, as there always is every time we hit a new thing. And I think a lot of, um, I think a lot of organizations did try to dip their toe into it. Now I've seen like, you know, as we're reading all the things that are out there, I've seen some crazy things like, like from like nobody, you know, successfully using them all the way to like 70 % of all existing organizations are now using AI agents, which I'm like, totally like BS, you know, it's just not, not true at all.

4:34It's a long way from the truth. Um, but I do think a lot of organizations are, uh, kind of whiplashed right now and kind of going, holy cow, what's this agent thing? I'm reading about it everywhere and we're trying to figure out what to do. Um, and that's happening at a moment of like, you know, where the, the, like those who are diving in and finding a use case that they can find success with, which is not easy in all cases, are making some big wows within their little world. And then those who aren't are still kind of fumbling in the dark. And I think that's fair. That doesn't mean that one person is smarter than the other.

5:15It just means that looking into the right use case and getting the right people to address it and having a good business case for it makes a lot of difference. And, you know, as we talk about that, I think one way of kind of leaping into that, you know, dichotomy is, is Andre Karpathy, you know, put out a post on a tweet on X. Do we still call them tweets? I don't, I don't. I have no idea. I'm not sure. But on X, and I won't read the whole thing because it was a fairly long one. But he's basically acknowledging, I mean, you're talking about this is one of the world's preeminent AI researchers that, you know, within our little AI bubble world, you know, on the technical side, he is a superstar in every possible way.

6:04And he's kind of saying, holy cow, and I'm totally paraphrasing, he didn't actually say holy cow. He's kind of saying, holy cow, even I at moments are feeling a bit left behind with how fast this is changing. And in the in the context, he's kind of talking about like coding and stuff is that, you know, after leading in the last few years and seeing models, you know, it seems quaint to talk about models, as you pointed out now, but talking about these models that are getting better and better steadily, but they still weren't doing great coding. You know, if you in terms of that, and the need for senior engineers to kind of correct it, and was it more trouble to use the model and the agent to do the coding or not?

6:48And did you just spend more time fixing errors? Well, all that really changed at the end of 2025. And with Opus 4.5 and OpenAI's 5.2 model in particular, as well as several others, but those are the ones that are called out the most, they got to where they could do senior level coding really well without mistakes. And I've griped because I'm a Rust programmer that because that's such a small community of programmers overall that the models weren't as good as they would be in Python and JavaScript. Well, guess what? It's kicking butt in Rust now. No longer Rusty. It's no longer Rusty. And so I, for one, as I am upskilling And as someone who has been using AI as we have gone forward in coding, it's now like my workflow has changed dramatically in the last two months in terms of understanding how to effectively use coding agents to do that.

7:54And I think we're, I don't think coding is the only area that that's impacting. I think there's a lot of areas where agentic AI, once you get a use case that is giving you some sense of success is like changing the, the, the field that, that small field that you're playing on in that. And that might be happening many times over. What do you think about? Yeah. Yeah. It's interesting just to read a little bit of that tweet that you reference. Karpathy mentions, clearly some powerful alien tool was handed around, except it comes with no manual and everyone has to figure out how to hold it and operate it.

8:36While the resulting magnitude nine earthquake is rocking the profession. And he kind of ends saying, roll up your sleeves to not fall behind. So, yeah, I definitely have felt this, Chris, just from my perspective. We get to kind of wax poetic on these episodes where it's just you and I. But from my perspective in building a company over this past year, Prediction Guard, I'm reflecting on our last board meeting. And the reflection back to us as leadership was, wow. Essentially making the note that product-wise, you all were able to advance so much more quickly without expanding your team in those last two quarters of the year as the company has progressed.

9:35So these are obviously investors, not that they have no technology, but they're not coders. But just from the output, the pace of development of the product and what we're able to achieve, it's significant enough to be noticed in that way without the larger team that would typically have been kind of required to, you know, reach that, you know, scale or support of what we're, you know, what we're supporting. Yeah, I want to relay a moment that I had. And as you know, like I sent you a text over the holidays, saying, holy cow, we got to talk about this at the beginning of the year and stuff. And I want to I want to share with you now because you I did not relay kind of like, what happened to me that made me send that text, because it's very relevant to this is I had been proposing a really complicated autonomy-based project for work.

10:39And I'm not going to get into specifics on what that is. But to take out the hype for a moment, I had spent a lot of time thinking and researching all the different things that had to go into that. And it was a tremendous amount of complexity involved in that and developed a really complex and very detailed and specific prompt on how to get there at like a production quality where it wasn't like what we would have talked about a year ago where it was like AI slop code coming out. And so I finally got to this point where I tried that out and like it in the matter of six minutes, it produced what I would have at least six weeks of work, at least six weeks of work in, you know, in just a matter of a handful of minutes.

11:31Now I had knew what I needed to put in. I knew a lot of that stuff and I was able to get a really good prompt going, but the actual work, like suddenly I had a large project laid out in VS code that had all these different things tied together. And I was just, I really, I just literally like, I don't think I've ever had that big of an aha moment, uh, in coding. And that was, and I was like, you know, tell Dan about this. So that was, I just wanted to, that was what prompted it. And it made me that like, I flipped over and realized this is the way forward and I'm 100 % in. So just to pick apart a little bit of what you said, Chris, there's some highlights there that I think are takeaways from our agentic work in 2025.

12:16One of those is with, I would say, No doubt at this point, these agentic workflows, especially driven by folks who have the relevant domain knowledge, are transformative in ways that are legitimately transformative, multiplicative, however you want to say that very much. I think we can confirm that. However, I think one of the things you highlighted is some of what was highlighted throughout the year around like the MIT study of things, you know, failing. Gartner says, you know, 11 % of organizations have agentic AI in production and that 40 % of projects will fail by 2027. I think part of that is driven maybe by two things that we've seen over this year.

13:08One is you do actually need to have a certain level of expertise to know both how to prompt and configure these systems, but also what data sources to connect into them, how to utilize this sort of, as Karpathy puts us, this alien tool, how to hold it, how to add in an MCP server. What type of automation am I really doing? How should it run? How do I integrate it into my day-to-day workflow? If you have that expertise around the integration side and infusing the domain knowledge, that is a key piece of it. And without that, there can be a lot of failure. Secondly, I think sometimes people are just trying to automate processes that are problematic because they're bad processes.

13:59not because the automation is bad, but they're just bad processes to begin with. So, you know, AI doesn't solve that problem. Yeah, there was one other takeaway that I'll throw in before we move on from this topic. And that is prior to developing the, I can't call it a stub for the thing because it was too much code. but prior to that kind of final prompt that got me well into the project I had there had been literally many hundreds of prompts before that which got me ready for that and I think a key thing that I came away from that with and which I've been sharing with other people since over the last few months is that that expertise is important for how you shape prompts to get the thing you need and you learn from it.

14:50So like at no point was the AI leaving me behind. Um, it would open up new doors, but I had to walk through those doors, take the learnings and develop the next prompt from it. And I think to your point a moment ago about kind of that expertise is it took that combination of domain expertise with how you prompt your way through a long sequence of prompts to finally get to the point where like you understood the system, well enough to where you could describe it in a prompt well enough so that a sophisticated agentic model could put the whole thing together in like nearly a production ready mode.

15:30So there was a lot of learning involved in that. So it wasn't just magic in five minutes. I was just quite taken by having gone through that long process, being able to do that final prompt and have so much produced that was at the quality level that I would have demanded it be. You know, for most developers, you've had this call, marketing calls, sales calls, and they want a new landing page. They want to redirect. They want designs implemented. And of course, engineering says, yeah, we'll get to it. But that bottleneck is why thousands of businesses from early stage startups to Fortune 500s are choosing to build their websites in Framer, where changes take minutes instead of days.

16:16So our friends at Framer, they're an enterprise-grade, no-code website builder that gives designers and marketers the ability to fully own your dot-com without having to rely on the engineering team. It works like your team's favorite design tool with real-time collaboration, a robust CMS with everything you need for great SEO and advanced analytics that include integrated A-B testing. Changes to your Framer site go live to the web in seconds with one click publish without help from engineering. That's priceless. That keeps you on task, on target delivering features. And that's how your team reduces dependencies and reaches escape velocity.

16:51And this isn't a toy. Framer is an enterprise solution with premium hosting, enterprise grade security, and 99.99 uptime SLAs. Companies like Perplexity, Miro, and Mixpanel trust Framer for their websites. Whether you want to launch a new site, test a few landing pages, or migrate your full.com, Framer has programs for startups, scale-ups, and large enterprises to make going from an idea to live site as easy and as fast as possible. Okay, so the next step is to learn how you can get more out of your.com from a Framer specialist or get started building for free today at framer.com slash practical AI for 30 % off a Framer Pro Annual Plan.

17:36That's framer.com slash practical AI for 30 % off framer.com slash practical AI. Rules and restrictions, it may apply.

17:50Chris, I think the other or at least one other theme that I know that we highlighted kind of going into this year was multimodal AI. I don't think, unless I'm misremembering or not seeing the right transcript, that we predicted kind of this reasoning era with models. So if we just look at the progression of models, which is definitely not the whole picture, as we just talked about, a lot of what happened was around agents, which for those that are listening, maybe new, you know, parsing through these terms, an agent is not just a model. It is a model that is connected to various external systems, some of which could be AI, some of which could not be AI and actually interacts with those systems to accomplish a goal.

18:44But so we're not just talking about models anymore. We're talking about these systems. But in terms of the models, I think we predicted more multimodality kind of coming into this year, which certainly we have. There have been many different vision language models, video models, music models, all sorts of things. Sora, all of these things that we've seen over 2025. five. And then there's these other reasoning models, starting with the multimodal ones, Chris, I think at least where I'm sitting, and it could be in just in my role, or the types of things that I'm seeing, but in the majority, actually, I think, yes, so I guess I would say all of the customer interactions that we are having and the people that I'm talking to really are using multimodal AI in terms of multimodal on the input side.

19:46I still very much do not interact with people that are doing kind of multimodal on the output side. So what I mean by that is certainly I see videos of coming out of Sora as reels on, you know, social media. And so I know that that is happening. Right. But in terms of the business world, real business context, I definitely see, you know, video, audio, image and text going into models, but not so much coming out really still coming out is either text or some form of text, like some structure, like a JSON structure, a tool call, some template, some, you know, field, some whatever it is, not really multimodal output.

20:38Maybe the exception to that might be synthesized speech, which is pretty pervasive everywhere as a thing in and of itself. So that's maybe a standout. I know I think you're right I like it that is a standout but I would have thought of that as a single mode like on the output side and I think you're calling out I think that there is a big opportunity here for especially uh when you combine it with agents in different capacities to have a richer output experience because at the end of the day I mean I know I know that um like Whereas my non-AI industry family members are more taken with the videos and things like that for entertainment.

21:22But with the work that I usually do, it's more that text output. And I could imagine a much richer output experience, to your point. It's easy to envision, especially when you think about, you know, like I'm going to dump all the different things into my input that I want it to process and assess for the output. but the output's still pretty basic. It might be that real-time voice interaction that was so hot six months ago, you know, that everybody got into and then it kind of passed and, you know, we all get our expectations set. You know, like, oh, okay, that's just real-time voice, no problem.

21:59But if you were to put a bunch of things together on output where you're getting text, you're getting that real-time voice in the conversational sense, you're getting supporting media, I think it could level up. So it probably will. Yeah. And I guess maybe one of the themes from this last year that we can take away as well is this rising up of the reasoning era. So reasoning models. We've talked about these on the show. Just as a reminder for people in case you missed out on our discussions for this year. In some ways, these reasoning models are mislabeled because they don't reason about anything.

22:41They just produce text. What is interesting is that they produce a segment of text that imitates or mimics reasoning or a chain of thought. Right. And kind of quote unquote thinks through a problem by generating text representing that thinking through of the problem. And then they generate a final answer, which has proven to kind of help pick through more complicated tasks, maybe do, you know, more orchestration or dynamic type of workflows than what we were seeing before. and these models I would say the many of the models that I see being released now at least in terms of that LLM flavor of models are either straight up reasoning models which means they're always going to reason in this way they're going to generate the reasoning tokens and then output the regular tokens or they are kind of conditionally reasoning models or hybrid models that We'll do that some of the time and not other of the time.

23:49And there's various implications of that. Certainly, I think you see that driving certain of these agentic, you know, working towards these agentic workflows. It also, to be honest, is sometimes annoying because often like in real business, you know, applications like we're working on, you really, if you dial in your workflow, you really don't want those reasoning tokens because they take so dang long, right? You have to wait. There's so much latency introduced by waiting for these reasoning tokens that unless you're doing this sort of very, very dynamic workflow, it's kind of annoying that a lot of these later models have these.

24:34I would say in general, it's a good thing. So we're definitely in the reasoning era and it's been cool to see these models come about but um you don't get anything for free uh there's a lot of latency that's developed and um because these models stream output right um every token that is generated is a inference run of the model meaning if you're generating 2 000 tokens of reasoning that's 2 000 runs of the model that is operating on a computer with a gpu that is expensive somewhere, right? You know, I agree with all that. But I think to some degree, it's intentionally or unintentionally, and probably the former rather than the latter, being driven by the organizations hosting these models.

25:25Because, you know, as one example that most, you know, that everybody would know is ChatGPT, you go in and you have a choice, you know, if I'm looking at their web interface, you have a choice between kind of instant or thinking. And of course, everybody wants thinking. Do you really want instantly to have thinking? And then if you choose thinking, then it's just standard thinking or extended thinking. And so that plays to a human bias of you're going to go for, well, yeah, I want it thinking and I want it extended thinking. And to your point, the cost of that may or may not be to you as the consumer, but certainly the cost of producing that is much more expensive with extended thinking on that, which kind of points out another thing that's happened over the last few months that we've all heard about, and that is that for years we talked about the limitation of having enough GPUs being the limiting factor on moving forward, and now it is power because you can take the same GPU and use it for many inferences, but each one of those separate inferences is taking a certain amount of power for that.

26:35So as a consumer of that, every prompt that I choose to make in a quote unquote reasoning fashion is going to be much more expensive in terms of power consumption. And we're hearing that in the news all the time these days. Yeah, I guess that takes us to an interesting theme that we've seen develop around infrastructure, hardware, energy. It's interesting to see that so much of this discussion, as you've mentioned, in recent trends, and I think this will continue into 2026, and it will create some both friction and opportunity and interesting dynamics in 2026, which is this limitation and opportunity around power.

27:21Just a couple of things anecdotally. I went to Colorado School of Mines as my undergrad, which, as the name indicates, still has a big tie to mining and petroleum and other things. And so I have friends in the energy industry and was talking with some of them how there's, you know, now very much speculators going around and trying to purchase and get the rights to power plants that were were relatively newly constructed but decommissioned while, you know, people were moving away from from coal. but speculating that these power plants will necessarily need to be turned back on. And other anecdotes like in our town here, Lafayette, West Lafayette, there's this huge, I forget, number of billions of dollars investment in a chip assembly plant here on the West Lafayette side.

28:35What's interesting to see all of the community back and forth to get the zoning approvals and the backlash that is happening against this chip assembly plan. And I'm not saying on one side or the other of that, but what I think is interesting is you see that dynamic here, right? In China, if you want to dominate in the AI space and you need a bunch of power plants, right? No city is going to say, no, we're not going to have our power plant here. They're just going to put a power plant there, right? And so that's how this has then filtered into this geopolitical space and environment that we're in where power and AI and chip manufacture and onshoring, all of this is what's driving now the political conversations.

29:28And so, yeah, we've seen this trend of from just having access to GPUs all the way kind of flow to these discussions around energy, infrastructure, power, which I'm sure will just continue throughout 2026. And it's and to delicately point at geopolitics and the implications, you know, some countries are now invading other countries and taking their oil. And, you know, that's, regardless of which side you're on, that was a notion that was kind of inconceivable. But power is the thing that people are talking about because every nation with its drive for more and more power consumption to support not only its normal things, but AI growth, as is the United States, you see a lot of interesting things happening there.

30:21So I'll just leave that one right there. I think, like to your point, this isn't a political show. We're talking about the practicalities of AI. But I think in thinking about the trends of 2025 into 2026, you can't go into 2026 without noting that when things happen politically across the world, AI is being mentioned as a motivation for why these things are happening. regardless again of who's doing right or wrong or your stance on something. We've moved from, I think, at the end of 2024 to now the end of 2025 going into 2026, where AI is the topic that is driving some of those policy decisions versus, I think, last year, if I was to kind of summarize, we were talking a lot about, well, how might AI or how might governments regulate AI as a kind of piece of their policy?

31:26Now it's almost driving the key pieces of policy in a lot of ways. Indeed. I guess going forward, it will be interesting as we go through 26 and see how policy continues to evolve in this, because this is a level of consumption that, you know, obviously is becoming a challenge to maintain and to even to initiate, because it's not stopping with where we're at, it's going on. So infrastructure, hardware, energy, those topics will, should be, it should be a volatile year in 26 to see where things go. Well, Chris, there's, of course, many, many things that have happened in 2025. And the majority of those we've talked about so far are related to Gen AI.

32:20I think in terms of practicality, moving into 2026, we would not be practical AI, I think, if we didn't highlight the fact that, you know, we recorded another episode right prior to this. and I won't give away anything that's in that episode other than there was one statement that said, hey, one trend that's happening with AI models is that Gen AI models have sort of plateaued on this transformer architecture that most all of these models are based on. But predictive models still continue to advance in a quite rapid pace. And what I mean by that for those listeners, again, that I'm kind of parsing through this jargon is these generative AI models like large language models, language vision models, et cetera, generate tokens or certain output like images or other things.

33:21Other models are discriminative or statistical and make predictions of classes or forecasts or those sorts of things. And the reality is that across industry, these models still continue to provide amazing ROI and get better and better. And the tooling actually gets better around those. And actually, what's interesting to me, Chris, is years ago, we talked about kind of this idea of auto ML, which is still a term that people use. There's still some things out there related to that. This idea that we could maybe automate the parameterization of AI or statistical models. And that would kind of help us create these models better and faster.