The AI Daily Brief: Artificial Intelligence News and Analysis - How AI Starts Doing the Work in 2026 With Anthropic CPO Mike Krieger

Episode Date: December 24, 2025

Anthropic CPO and Instagram co-founder Mike Krieger joins the AI Daily Brief to talk about the rise of vibe coding, why coding agents quietly became the breakout AI use case of 2025, and how enterprises are beginning to move from chatbots to real workload-taking agents. The conversation explores how tools like Claude Code escaped the developer box, what it takes to design products for capabilities that don’t fully exist yet, and why 2026 may be the year AI starts reliably taking work off people’s plates inside large organizations at Anthropic. Brought to you by:KPMG – Discover how AI is transforming possibility into reality. Tune into the new KPMG 'You Can with AI' podcast and unlock insights that will inform smarter decisions inside your enterprise. Listen now and start shaping your future with every episode. ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.kpmg.us/AIpodcasts⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Blitzy.com - Go to ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://blitzy.com/⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ to build enterprise software in days, not months Robots & Pencils - Cloud-native AI solutions that power results ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://robotsandpencils.com/⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠The Agent Readiness Audit from Superintelligent - Go to ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://besuper.ai/ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠to request your company's agent readiness score.The AI Daily Brief helps you understand the most important news and discussions in AI. Subscribe to the podcast version of The AI Daily Brief wherever you listen: https://pod.link/1680633614Interested in sponsoring the show? [e-mail]

Today on the AI Daily Brief, The Future of Vibe Coding and what's in store with AI 2026 with Mike Krieger, the chief product officer of Anthropic. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, super intelligent, robots and pencils and blitzie. To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can subscribe on Apple Podcasts. To learn more about sponsoring the show, send us a note at Sponsors at AIDlybrief.A.I.D.B.I.I.: and if you are interested in our forthcoming AIDB Intelligence Service, check it out at AIDBintel.com. Now, as we move forward into our end-of-year episodes, I'm excited to add a couple of conversations into the mix.

You might know Mike Krieger as the co-founder of Instagram. Real Ones will also know him as the co-founder of Artifact, an AI-powered news app. However, for most of you right now, Mike's most important role is as the chief product officer of Anthropic. In this conversation, we talk about the origins of Anthropics focus on coding, how enterprise AI uses has changed over the course of the year, and some of the trends that Mike is most excited about heading into 2026. All right, Mike, welcome to the AI Daily Brief. Great to have you here. It's great to be here. Thanks for having me. Yeah. So this is super fun, like I was just saying, some of my favorite episodes of the year, or these end of year episodes where we get to kind of think big, look forward. And one of the

big themes, I think, for me heading into the new year is sort of everything vibe coding, everything agentic. And so I was super excited to have you join the show. What I wanted to do, though, is actually kind of go way back a little bit. I think, you know, a lot of folks see Anthropic as sort of the torchbearer in a lot of ways for AI coding. And I wondered, you know, I was thinking about when you join the organization and just how, how early was that sort of focus clear? You know, was that an emergent phenomenon as it became clear that there was something very differentiated in these models and that's sort of how people were using it, or is that sort of like intention from very early on, that this is a broad sort of set of use cases that matter to you guys. Yeah, the thing I always

like to say, whenever there's sort of product folks inside Anthropic, they're thinking about sort of which direction to take things in is the more you can align with the sort of company's general long-term perspective about where powerful AI will come from, like the smoother things will go because Anthropic is nothing but focused, right? And I think that that's shown through in sort of the, you know, the bets that we choose to make versus not. And definitely there's this belief that for very powerful AI, you need the ability of the model to reason about things, to plan genetically and work for a long time horizon, but also to be able to write and run code, not only to produce software, but because it's a really useful tool for solving problems.

And so that belief was in there and it predates me. I joined in May of last year. But it kind of coincided sort of with the outside world realizing it, because Cloud 3, which had come out, I think, a month before that, was the first model. And I remember there was like that moment on Twitter, and everybody said, oh, wow, this model can actually write, not just like sort of function level, but like entire, you know, sort of files of code. And of course, compared to now, it was not very good at it,

but it was already, you know, amazing what it could do then. And then we paired it with our first sort of more coding oriented product, which was artifacts. So you could have, you know, cloud kind of generate, you know, at the time was mostly React sites, you know, alongside the chat. And that was kind of, I think, for a lot of people, the first moment they realized,

oh, this is an interesting new experience of kind of coding alongside the model and not necessarily doing it in a development environment. Yeah, it's interesting. I think you can in a lot of ways almost chart people's sort of the viability of a lot of this two key releases alongside Anthropic. You know, I remember when I first started this show, it was actually April of 2023, and already sort of agentic coding was like the thing that people were most excited about, like a GPT engineer, which would later actually become sort of morph into lovable

like 18 months later or something like that, was, it was like my first viral YouTube episode was about GPT engineer. And so it's interesting to see kind of like at each stage how more use cases get unlocked and sort of a broader set of people come into the fold. Coming into 2025,

you know, I think that the odds-on favorite for what the year was going to be about, at least if you looked back at kind of all the AI content creators, it's going to be the year of agents, right? And I think looking back, it was, but it was the year of coding agents. Did you guys have a sense coming into this year that this was poised to be kind of, you know, the significant use case or the breakout based on what conversations you were having based on the capabilities that you were seeing? Yeah, it's a great sort of moment to reflect because going into the sort of last couple weeks of the year, last year, we had built something internally we called Claude CLY, which we later released as Claude Code. And the emergence of that came from our labs team,

which is a team that really focuses on trying to do sort of disruptive zero to one ideas. And that was everything from like early computer use explorations and some wacky things. And also this Claude CLI thing. And between, I think, September, when the first version got sort of rolled out internally to December, it rapidly overtook every other sort of coding tool we had internally.

And it was because it kind of had this bet that the models are going to be able to do more and more, maybe not this model, but the next one and the next one, the next one. So let's let the model cook for a longer. Let's let it sort of act for, you know, longer periods of time. And so that way, you know, going to the holidays, it was that question of, do we release this? You know, like, do we now add a, you know, kind of third component to the product portfolio beyond just cloud AI and then and the API? And so that was the active conversation that was happening. But we really felt like, if not us, then at least somebody using our models would sort of co-discover this piece where you don't need to hold the models so closely anymore.

you can let it operate over a sort of fuzzier task definition and over a longer time. I think it still needed a fair amount of handholding then, but you could see, you can see the shape of it. So it was definitely the coming into this year, we felt like that was going to be a major shift in how people were going to build software. Well, you know, it's a super interesting quang. One of the things, you're sort of, you know, you have a deep product experience. And one of the challenges I think now for product folks and just for entrepreneurs in

general is there's this sense that to be successful, you have to, you know, not just gives lip service to the idea of skating to where the puck is going, but actually sort of design and orient what you're building for capabilities that do not yet exist. And that's an extraordinarily hard thing to do. And it sounds like that was part of the genesis of Claude was just some sort of attempt or, you know, some like, you know, scratching against that itch in some way. Yeah, we have product principles inside and throw. And one of them is write the exponential, which is like we're trying to build products that both, you know, meet the moment. So they're useful today, or at least they poke at something useful today, maybe the ones that are a little early, we won't release yet, but that they can naturally improve.

And it's been interesting. Even on the cloud code side, we've deleted parts of the harness over time rather than added to it because the model can do more. And it's really interesting also. We work with a lot of kind of downstream customers that are using the model. And sometimes, you know, we'll drop a new model, you know, a research model. And it doesn't look like it improve very much. And then we'll send some apply to I, you know, folks to spend time with them. They realize, right, now we're actually harness bound.

And we need to actually let them evolve and let the model do a bit more to, like, loosen that as well. But it's definitely an active conversation that we have with folks building on top of the platform. Like, you know, they have some visibility about where we're going. Like maybe they'll be in a research program early access. But they still have to do a fair amount of this. All right. So if the models are here now and I need to do this much additional scaffolding, what does this look like if I need to do less scaffolding?

is my product still useful in adding value? And can the model then do even more for me? Or is it now going to squeeze the piece that I thought I was adding value in? Have you been surprised at all with the way that people have used cloud code since you released it? Because it, you know, it is much broader uptake than just sort of, you know, core audience of software engineers. Yeah, absolutely. We, internally, we, you know, had this internal project that people were using and then we, like, buttoned it up and put on a more fancy suit to be able to release it publicly. But then, as you can imagine, like, the internal use cases kind of kept co-developing.

And so we do like every two to three times a year, we do a hackathon. And it's been notable that every hackathon we've done has been around the time that some technology is like poised for a breakout. So the first one we did was around MCP. And every single project was MCP-based before, really, the rest of the world had kind of caught onto MCP. The second one we did was around the time that Cloud Code had been released. And what was really interesting was how many projects were not coding projects,

but they were using CloudCode as the underlying engine. So there was one that was using CloudCode. for doing bioinformatics, which we later kind of channeled into Cloud for Life Sciences. Another one that was using Claude as a sort of SRE in a box and was able to use Cloud Code as a way of, you know, looking at data sources. There was Cloud as a data scientist. There's all these sort of pop-up projects that was nice that they didn't have to reinvent the tool used kind of bit that could just add value on top of that. And then when we launched it, we started seeing things externally, too, like people using Cloud Code as their project manager,

Cloud Code is their PM, Cloud Code as a data sciences externally. So you started seeing this much more. It's why we eventually renamed the underlying SDK to the cloud agent SDK, because we realized calling it code was doing it a disturbance relative to what kind of use cases we were actually seeing. Yeah. So this is one of the questions that I'm most interested to see in coming year, but even the coming years, is what it takes to kind of rewire people with these new tool set. It's like this whole language, this whole infrastructure that they have access to, especially if they, that they haven't before, especially if they're not developers. Do you think that some of this, you know, if on the spectrum from this early kind of usage of

Claude code for non-coding use cases is tinkerers who are kind of, you know, more technical than they let on the one end of the spectrum versus actually kind of heralding a different set of interaction patterns that people are going to have, how do you see that evolving? Yeah, I think it's early still. Like even when we look inside companies that have deployed like Cloud for Enterprise and they have builders within their sales team or their ads team or whatever, different non-technical team, you will always find this sort of persona, which is the tinker builder, like early adopter within that space that usually is not an engineer.

It doesn't even have an engineering background, but has figured out enough and as like learned the primitives and can then talk to cloud enough about how to fix these issues, that they can then kind of build something pretty powerful, whether it's, you know, automating part of what they were doing, sort of enriching what they were doing, making their team's lives. These are like all these different pieces, but it does still take that person, which I think is probably a natural part of the kind of software life cycle. We are. I think there's still this gap.

And I think that's both a gap in interface in terms of how people think to interact with and how these products reveal their full capabilities. And then also the actual capabilities themselves where if you had a human coworker and it was very creative at solving problems, like you gave it a high level task, I was able to do it most of the time. but sometimes it would sort of make a mistake that you would never have expected to make. Based on it having just done it great last week, you'd be like, have a pretty complicated relationship with that coworker. I still were at that phase still of this like gap between understandability of these systems, but then also gap of, you know, how reliable and predictable are they when they do start working? And can they feel more like a thing that gets just predictably better over time?

Yeah, I think that's true. And I also think that there's just, you know, I don't have the exact right words for this, but there's some lag in terms of just, you know, unwinding and undoing however many years or decades of the way that you've been doing a thing before that is, it just takes time. You know, I think about it, you know, we're now, I guess, 10 months into vibe coding as a named phenomenon, right? It was the same February of this year, the same month that Cloud Code came out. And I'm still finding myself as someone who literally, you know, podcasts about this every day and is living inside these tools. I'm only just now starting to find myself actively ask on a regular basis, like, could I be building something to do this instead of using a Google sheet or instead of, you know, however I used to do it? And again, that's, that's me as someone who's as deep in this as you can get.

I think there's something really to the, you know, building with one tool that gets you comfortable with it and familiar with it. It's easier to build the incremental n plus one, but it's that first one that requires that sort of uplift if you're not in the habit. So I was working on a project over the weekend. I was using Replit and using Opus under the hood. And then I also needed to create a secret Santa for my family. And that, you know, because I had been in the tools already, it was over breakfast while I was cooking eggs. I kind of kicked off this asynchronous request.

And by the time I was done, it actually had built the whole thing. And that was really cool. But I wouldn't have reached for it as my first tool had I just not been, you know, sort of interacting with that same software. So I do think that there's this sort of still like habit creation and adaptation of even knowing you can do that, that we still need to close. Yep, exactly. Sort of similar example.

Someone had mentioned building a gift tracker with one of these tools. And I love that because we always end up. It's Christmas Eve and I'm like, we're shoving presents in the closet for later because we just bought too many things. And so I copied that. I imitated it. And it was probably for whatever reason the first time in a month that I built something. And the just huge, basically since this last wave of, of sort of,

sort of models had come out. And within, I don't know, three or four days, I had like six or seven different applications that it just sort of like, you know, spiraled from there. I think it's also interesting. Like when we, you've seen this not even just for our model launches, but other model launches, there's been the sense from people that are primarily maybe interacting with it, you know, in chat. And I'll say, oh, it seems a little smarter. Maybe it's a little bit warmer. But they're very sort of vibes based assessments. And if you're not sort of checking in and dipping back in and trying to build something with it, that's where I think you see the biggest leaps.

I definitely saw it over the weekend actually getting to build with Opus, for real, over two days of getting really deep in there. I was like, oh, these are applications that would not have been doable, even in Sonnet 4.5. It would have hit some ceiling or it would have gotten stuck in some loop, but then just watching it hit a wall, you know, debug itself, tail the logs, add debug logging, roll it out again, pull up a browser, check the, like all of these capabilities

that I think have just either been co-created along with the model or now the model is using better as it's improved. But you really got to sort of dip in and push it to really see that difference. But I think where there's some of that jadedness, I think from the outside sometimes of, oh, are we hitting a plateau? Whereas, you know, if you checkpoint it, you can definitely see that continued improvement. Today's episode is brought to you by my company, Super Intelligent. Superintelligent is an AI planning platform.

And right now, as we head into 26, the big theme that we're seeing among the enterprises that we work with is a real determination to make 20, 2026 a year of scaled AI deployments, not just more pilots and experiments. However, many of our partners are stuck on some AI plateau. It might be issues of governance. It might be issues of data readiness. It might be issues of process mapping. Whatever the case, we're launching a new type of assessment called Plateau breaker

that, as you probably guess from that name, is about breaking through AI plateaus. We'll deploy voice agents to collect information and diagnose what the real bottlenecks are that are keeping you on that plateau. From there, we put together a blueprint and an action plan that helps you move right through that plateau into full-scale deployment and real ROI. If you're interested in learning more about Plateaubreaker, shoot us a note, contact at B-Super.a.I

with plateau in the subject line. Small, nimble teams beat bloated consulting every time. Robots and pencils partners with organizations on intelligent cloud-native systems powered by AI. They cover human needs, design AI solutions, and cut-through complexity to deliver meaningful impact without the layers of bureaucracy. As an AWS-certified partner, Robots & Pencils combines the reach of a large firm with the focus of a trusted partner. With teams across the U.S., Canada, Europe, and Latin America, clients gain

local expertise and global scale. As AI evolves, they ensure you keep peace with change. And that means faster results, measurable outcomes, and a partnership built to last. The right partner makes progress inevitable. Partner with Robots and Pencils at Robots and Pencils.com slash AI Daily Brief. This episode is brought to you by Blitzy, the Enterprise Autonomous Software Development Platform with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand Enterprise-scale

code bases with millions of lines of code. Enterprise engineering leaders start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and pre-compiles code for each task. Blitzy delivers 80% plus of the development work autonomously, while providing a guide for the final 20% of human development work required to complete the sprint. Public companies are achieving a 5x engineering velocity increase when incorporating Blitzie as their pre-IDE development tool, pairing it with their coding pilot of choice to bring an

AI-native SDLC into their org. Visit blitzy.com and press get a demo to learn how Blitzy transforms your SDLC from AI-assisted to AI-native. Well, so this is actually an interesting point at which to sort of maybe try to fork the AI coding or vibe-coding conversation into a few different buckets. You've got sophisticated, the sophisticated sort of software engineering conversation around what AI coding is going to mean and sort of, you know, the autonomy spectrum and, and sort of where these models are and what they can do. And you almost have, you've gone through kind of a full cycle this year where, you know, it was, you know, huge amounts of uptake, but now we're kind of sifting through the new challenges, right? You know, new technology doesn't solve all problems. It trades one set of problems for a

hopefully better set of problems, but then you still have to solve those. But then on the other end of the spectrum, you've got, it's incredibly nascent, the individual, the non-technical usage of these tools. I think we've barely started to scratch the surface on those things. And then somewhere in the middle, you've got kind of enterprise usage, which includes some of that software engineering reorgan organization, but also includes, I think, lots of folks who are thinking about how to use these types of tools for other aspects of the business rather than just sort of, you know, building software. How much do you think these are the same conversation versus, you know, again, three, three three kind of or two or three different conversations using all the same words.

Yeah, no, I think you're right. I think even if they have an underlying model, that's the same. And even if some of the other building blocks that you might use in there, either from an SDK perspective, they all sort of end up needing different applications or a different sort of manifestations. They feel quite different. I think you're right. So on the software development side, you have in general, like a pretty motivated population

that has always been interested in tinkering with their tools, like hence the Emacs versus VIM, like, you know, tabs versus spaces. is like programmers maybe notoriously have the desire to sort of optimize their own building environment, which other disciplines might have, but not always to the same degree. And it's not always as easy to swap one thing out for another. So the adoption there, the evolution has become, I think, this flywheel where it's really clear for, you know, engineers in Anthropic, where the model needs to improve. And that's very helpful for us to coordinate with research and close that flywheel as well

as, you know, feedback from external folks. In the middle piece, there's still this sort of ceiling of complexity that you can hit. Now, there's been really impressive sort of zero to one, or vibe-coded applications that have even been released, but you still, I think the gap I was perceiving, even just watching my wife, who's a sort of product manager, UX designer, by training,

not a software engineer, use some of these, is that you still sometimes need to know the right sort of magic incantation words to use. We were building a product together, like a side project, and the way it was using LLMs was just filling the context window, and I was like, okay, well, you actually probably need to move to some, like, semantic retrieval piece, but Opus wasn't suggesting that out of the box,

and she didn't know the magic words. And it took me saying, all right, we actually probably need to move to this embedding solution. It was just like a layer of complexity above. And so I think one thing that our models, all of these coding models can do a better job of in that middle category of helping non-technical people build things that are effectively software,

is helping them move up that complexity ladder in a sort of more structured or thoughtful way, where, yes, the five-coded front-end only thing is great to show off an idea, and then you want to persisting, Okay, that's the next step. Or now you want to persist data. You're thinking about launching this,

that's going to take a whole other step of security reviews and thinking about things. And like, oh, now you've launched it. And the thing is melting under, you know, load. Okay, great. Now I got to put on my performance engineering hat and then build from there. And the same way, you know, with Instagram, we went through that process of, first we were just building the UI. Then we built the back end.

Then we launched it totally fell over because it got attention. And then we kind of rebuilt it over the next, you know, weeks and months to sort of manage it. You've got to speed run that, but with model assistance now. So that feels like the. big piece on the middle one. And then the last one, I think on the, on the enterprise software side, you know, I saw you cover the, you know, like famous MIT report of that like gap of expectations. And I think that was such a, it was one of those things that was truthy. And even if there was like,

sort of methodological problems underneath the study, it did point to something that a lot of people had, which is like, I got AI rolled out to me at work, but I'm not sure I'm more productive. And I think the, the place to close that gap, I think there's a bunch of things. One is just making sure the output quality is actually good enough where you're saving yourself time, where something is half done, I think for most people, they said, well, I probably would have been faster doing this myself rather than hanging up with something that's like not quite there, and then I'm struggling to get it to iterate with me to where I will need it to be.

So a lot of the emphasis we've been doing is actually less on the agenic side. It's less like take my two-sentence description and generate an entire PowerPoint deck out of it. And it's much more, you know, require a little bit more upfront work, but really focus on making sure that that initial, you know, sort of thing that got created was high quality enough where you felt relief and happy that it saved you time rather than, oh, man, I've just created more work for myself by using AI. As you look into 2026, where do you see enterprises starting this year? Maybe especially as compared to where they were starting in 25, what do you think the big goals are that you have sort of thinking about, you know, both model design but also product design?

I think maybe two things that feel markedly different now versus a year ago. One is enterprise is getting more interested in rolling out. We've been calling horizontal agents, but basically, you know, if you think about the sort of companion agent or co-pilot agent where there's a strong human in the loop and you're kind of co-creating either a document or an email or, you know, whatever that may be, seeing also a lot more interest now in, great, we have, you know, this repetitive back office task. We're trying to scale up to, you know, handle international, know your customer requests, whatever those sort of complicated but repeatable processes that have something that is bespoke to that enterprise, but also something that is sort of regular. for example. We're seeing a shift there where there's a lot more interest, and we've been deploying, you know, applied AI and engineers into these enterprises to help them get those agents running. It's often about sort of translating what those requirements are into that, you know, process, again, where the model can be creative and flexible, but still repeatable enough to, you know, follow their operating procedures. That feels markedly, a year ago, we weren't

really having really any of those conversations as well. And the second piece, which also feels nascent is, I think all of these enterprises, especially any that have this public-facing product that they might be shipping, is kind of going beyond V1, which was like, let's kind of sprinkle AI on these different surfaces and hope that improves the product, too, do we need to rethink some fundamental pieces of the product to be more, you know, ancient native to use a buzzword? But what I really see it as is, you know, have you unlocked the full power of your product to any AI that is sort of running on top or alongside it? And we could talk about that.

But I think that's a harder transition to make than, right, now we've got a sidebar that you can chat with your, you know, AI and kind of integrate with the rest of the product. Yeah, it's interesting. I think that, again, kind of looking back at maybe what expectations were versus what played out, again, if we take the idea that 25 was the year of agents, but maybe a little bit differently than we thought, you know, one part of that was it was the year of coding agents. But another part of that was it was also the year of agent infrastructure. You know, this is a year where MCP became ubiquitous. more recently, I just today, before recording this, did a show about OpenAI adding skills support or starting to experiment with skills support. And it's very clear that everyone is much more interested at this stage at sort of building the necessary infrastructure to be able to move forward faster than in sort of getting waylaid in the sort of standards wars that we've had in the past. And I wonder, it feels to me like we're poised a little bit for enterprises to almost go

through their kind of infrastructure year in 26 where, you know, again, going back to the lesson at the MIT study or the truthiness of it, hold aside the specifics. The fact that it had such resonance suggests, and I agree with this, that there is something, you know, some gap there. I think that a lot of organizations are embracing now that it's just, you're not just going to drop a chatbot in, or you're going to do that, but then to really go kind of to the next level, it's going to involve a much more sort of, you know, comprehensive review of how you do things. And it feels to me like, you know, perhaps that some amount of that process redesign is what organizations, at least the ones who are kind of, you know, ahead are going to be in four and 26. Absolutely. And I was talking to somebody

who runs technology at a large bank. And he was telling me that they had to rethink, not just the data storage piece, which they'd already been doing a lot of work on, but also the sort of data annotation and sort of lineage piece to be more. more AI-friendly. So that when you asked, you know, Cloud to, hey, help me construct, you know, a dashboard on this or help me understand this data query, even having that additional layer of annotation or understanding of what these different tables are and what they represent, but a huge way to actually making that a useful sort of product.

And so figuring out what are the missing connector bits is going to be, I think, a lot of 2026, which is great. We have MCPs. We're seeing more and more enterprises, wrap some of their internal services or internal data stores and MCP so they can. get access to it inside, for example, Claude. Now the next turn is, that's maybe on the retrieval side. Can you actually start taking action and making it a useful

participant in business processes by enabling it to either, make a human assistant decision or make a cue up a decision that a human can conform, whatever the right sort of metaphor, human in the loop piece is? But moving up that complexity ladder, so that again, it can actually start providing value that befits its level in the discourse. I want to talk in our last few minutes about some of your predictions or thoughts about how 26 is going to end up differing from 25 with AI.

And maybe just to get us started, we were just kind of talking about expanded enterprise use cases. But what do you think are going to be the biggest blockers for enterprises and how do you think they're going to get through them? I think for a lot of enterprises that we talk to, there's still this gap between sort of the idealized, like, great, if you ran this perfectly on this like one cloud with all, you know, your, you know, permissions all perfectly. set and you're okay with inference happening in this way, then we could unlock use case tomorrow on the reality, which is there's legacy systems, there's often sort of regulatory reasons why they, for example, will only run in this particular way on AWS in this particular kind of setup. And so a lot of the work that we were doing for next year is the word we've been using is distributability, which I think the spell corrector tells me is not really a word, but what we

really mean is if we want to bring our intelligence and even our agentic primitives, you know, whether it's skills, whether it's the agent SDK, whether it's storage, whether it's memory, all of these pieces into actual enterprise workloads. We need to really actually embed and meet them where they are. And so there's a lot more work on, hey, let's actually like componentize this, make it available everywhere. You see it now that we're on all three major clouds. Like that, the general set of projects is kind of closing those gaps because there is

interest, and especially from the sort of more sort of forward-looking CTOs and CIOs, but they also do need to work. with sort of the existing constraints instead of they have. And you can kind of get the pilots done in a like pretty, you know, rough and ready way just to prove it out. But to really reach that production scale, I think that's the biggest blocker. Tool versus colleague.

This is something that we've been sort of talking about for a while. And I think this is maybe a false binary in terms of what, you know, when we reach maturity of AI. But do you think that we'll start to see more of that kind of treating AI, not just as a tool, but as a thing that can take on ever bigger workloads? Do you think that that sort of starts to come to reality next year? Yeah, I think that probably more than anything is what we'll define the year is you start seeing this already with coding. So we did this, get a partnership with their agent HQ piece.

We're now you tag cloud in a pull request and then you go have your coffee and you come back and it's done whatever you needed to do. And we did the same integration with cloud code. That's sort of pointing at the kind of interaction that you might expect. Now, is it already going to be at the place where it can onboard onto the organization, and understand the problem space, understand the sort of dynamics of all the relationships and just pick up work? No, I don't think we're going to be there.

Maybe near the end of the year, we'll have some kind of early glimmers of there. But I do think the sort of more piece of the job function that has like a clean sort of, you know, right, great, we need to prepare this kind of report. Here's the work I've done already. Here's where you can go get more information. Here's what good looks like.

Report back to me, you know, in a way that you might delegate to somebody else. That's very much around the corner. how we're thinking about a lot of our product strategy and coming into next year is, how do we enable that? What are the interfaces that we need to create that make that possible? And then what do we learn about what's working on the software domain that we can apply to knowledge work? Maybe asking you too much to put on a marketing hat. But if you add sort of a phrase for capturing, you know,

what you hope AI does in 26? What would it be? It's reliably take work off your plate. I like it. All right, Mike. Well, this is a super, super fun conversation. You could go for another half hour, hour easy. But appreciate you making the time.

And really excited to see where you guys cook up. It was great to view. Thanks for having me.