Did OpenAI Just Kill a Bunch of Agent Startups? OpenAI DevDay 2025 First Reactions

0:00Today, we are doing a rapid reaction post to OpenAI Dev Day, which is ongoing as I'm recording this and asking a bunch of questions with one right at the top of the list. Did OpenAI just kill a bunch of agent builder startups?

0:18All right, friends, we are trying something a little bit different today. AI has got to be the only field in the world right now where even with a daily podcast, there are some days when I am behind even with publishing something every 24 hours. Luckily, this time I knew that Dev Day was coming up, and so I wanted to use it as a way to try something new. Tomorrow, I will be doing the full deep dive on everything that was announced, plus getting into this big OpenAI AMD deal, which honestly has been talked about as much as anything at DevDay, if not more. Plus, we'll have the benefit of an overnight of people's reactions and playing around with what has been released.

0:50But I wanted to do a quick bonus rapid reactions type of episode to see if that's the type of thing that in the future you would like me to do. So this episode is going to be much more opinionated and just give you my first takes after having just watched the OpenAI keynote. If this is something you like, let me know in the comments. My guess is for all of us insane people in AI, more content is more content. But if for some reason you hate it, let me know that as well. All right, so going into Dev Day, we had started to get some rumors about what was coming. Some of it was more hopes and dreams than anything else, but some of it seemed solid.

1:23Particularly last night, it seemed like it was pretty clear that we were getting some version of an agent builder. So first, let's talk about what we did not get. We didn't get a new image model. In fact, we didn't get any new models in general, which is not unsurprising. Dev Day in the past hasn't been used as a place for new models either. It's really been about tooling around the models that makes them more usable, and certainly that's a lot of what we got as well. What we did get was first, a quick update on the status of things, 4 million developers building with OpenAI, 800 million weekly ChatGPT users, 6 billion tokens per minute on the API platform.

1:57Separately, we also heard that Codex had served 40 trillion tokens since its release. Now, in terms of big announcements, Sam Altman and the team at OpenAI bunched them into four categories. The first was apps and chat GPT. The second was AgentKit. The third was some updates for Codex. And the fifth was API updates. The two headliners were definitely apps and chat GPT and AgentKit. Apps and chat GPT is what it sounds like. It's a new way to interact with native applications that are built directly into ChatGPT. Users can call up specific applications by tagging them in, but ChatGPT can also recommend applications if they might be useful based on what the user is querying.

2:37They say that there will also be other methods of discovery in the future as well. Apps are displayed in line, they can render anything that they could on the web, and they also support picture-in-picture and expand to full screen. If there's a video, they're going to pin it to the top of the screen so you can still access your conversation as it's happening. By way of example, they showed an app from education company Coursera with one of the cool interactions that they showed being that you could pause the video that you were watching from Coursera and ask ChatGPT, can you explain more about what they're saying right now?

3:05The app's SDK, which is the formal name for what they released, has this feature that they call Talking to Apps that allows ChatGPT to have the context of what you're experiencing in the app so that it can answer that question. In other words, it is a really deep integration between the apps and ChatGPT. They aren't just sitting there in the ChatGPT window. ChatGPT is actually able to interact with them in meaningful ways that allow you to bring that assistant type experience to your consumption of the app. Another example they gave like that was interacting with a Zillow app, where after clicking on a specific house, they could also ask questions from ChatGPT that Zillow didn't have the information for.

3:45So the one in the live demo was asking how close a particular house was to a dog park. They also showed off an app from Canva, and additionally, Booking.com, Expedia, Figma, and Spotify all have apps available today as launch partners, with a whole bunch of others including Khan Academy, Instacart, Uber Thumbtack, TripAdvisors, and more coming soon. The app's SDK is, of course, built on the Model Context Protocol, or MCP. The second headliner announcement was AgentKit, and this is the one that people had started talking about the night before. AgentKit included a number of components, including an agent builder, which is their visual canvas for creating multi-agent workflows, a chat kit, which is a tool for embedding chat experiences into products and agents, a native evals platform that has everything from trace grading to grade agent decision-making step-by-step to automated prompt optimization, and of course, you can also connect other data sources to agents via OpenAI's existing connectors platform.

4:37They did a live demo where they took eight minutes to build and ship an agent right there. It was a simple little agent that turned the DevDay website into an interactive service where someone could ask it, for example, what sessions they could go to to learn about agent building. For Codex, they talked about a new Slack integration and new enterprise controls. And on the API front, they had a bunch of announcements, including GPT-5 Pro in the API, and a little unexpectedly for some at least, Sora 2 and Sora 2 Pro coming to the API as well. And by the way, I had said that they weren't highlighting any new model announcements.

5:10They did actually, as part of the broader Dev Day, have two little model updates, a smaller voice model called GPT Real-Time Mini that is 70 % less expensive than the larger model, as well as a GPT Image 1 Mini that's 80 % less expensive than the larger version. Those just weren't highlighted in the keynote. Okay, so that's the quick overview. So let's talk about the vibes and then answer a few questions. Overall, Allie Miller was in the room and said that in order of developer excitement, it was agents, codecs, and apps ranked highly scientifically by the energy in the room when OpenAI was talking about those things, the amount of phones out, the applause count in volume, and how much heads were moving to whisper to each other during the presentations.

5:47So the number one question, which had started even last night as soon as people started seeing that OpenAI was going to release some sort of agent builder, was did OpenAI just kill a bunch of agent startups. The ones mentioned most often were Lindy, N8N, and Zapier. So did OpenAI just kill a bunch of agent startups? The argument for yes is, of course, that OpenAI has such incredible distribution that competing against that is just going to be enormously difficult. It is already incredibly hard to build and maintain a moat of any kind when it comes to software, and so going up against the intense power of OpenAI just seems like a big task.

6:24Lindy struck a defiant, optimistic tone with Founder Flow writing, Welcome to the club, OpenAI, and posting a visual note that said, Welcome to the most exciting category in AI, and congratulations on your first foray into true AI employees. Zapier got a little bit more specific about why they think they're in a different place. They tweeted, Agent Builder was just announced, a new way to design AI-powered workflows right inside OpenAI, but it ships with only a few native integrations and most businesses run on hundreds of tools. So basically what they're arguing is that their ecosystem of 8 ,000 apps, 30 ,000 actions, provides something fundamentally different and frankly complementary to OpenAI's agent builder as well.

7:02I think a couple things. First of all, it is absolutely true that going up against something that OpenAI perceives as core functionality for their platform is not a super fun prospect. There's just no way around that. And especially given that they built this on MCP and seem willing to reach outside their ecosystem to be the place where it all happens, they are going to be a formidable competitor. At the same time, I don't think that these companies are wrong to recognize that there are going to be a huge number of people and enterprises that prioritize model flexibility. They're going to want to be able to shift in and out different models, not just based on them changing and improving over time, but for different use cases.

7:39It is inherently a limitation of any agentic experience that comes from a foundation model company that it will be limited to that company's models. There absolutely is a wedge there. It's just a question of how big a wedge. Another source of optimism is that the current visual workflow design type of user experience that has been used for a while now by these companies, Zapier, Lindy, and N8N, and is now coming to Agent Builder, is at the moment an extremely niche and frankly, fairly intimidating user experience. As much as these companies say that they're not for technical people, there is still a big hurdle there.

8:14It is entirely possible to me that having a major company like OpenAI offering this helps normalize that user experience and maybe expands overall business for this type of agentic workflow builder. The other part of this is that the way that OpenAI is coming at this is, at least for now, still definitely more technical than these other companies that we've just mentioned. Even the demo involved a lot of coding, and it's clear that at least at this point, AgentKit is definitely imagined and being designed as a tool for developers to integrate and build agents more quickly, as opposed to being a general consumer agent builder.

8:49Ethan Malek writes, early agent builder impressions are that it is very solid and a huge expansion of who can create agents, but at this stage may still be too technical and single player to be a true replacement for the dream of GPTs, where anyone might easily share prompts and use cases with teams and firms. Basically, as AgentKit has been shared and designed so far, it is unlikely to create some mass democratization of AI agents. So as to the question of did OpenAI just kill a bunch of agent startups, I certainly think that in some ways it made their life harder, but there could also be unanticipated benefits as well, and I don't think anything about the competitive set is a foregone conclusion just yet.

9:25Next up, is apps just GPT's 2.0? Now, it would be going too far to call GPT's a total failure. There are still plenty of people who use these tools and super intelligent. We have a couple of GPTs that we've built personally that we use, but they certainly weren't the breakout mass app store for AI kind of hit that some thought they were going to be when they were first announced. And so I think for some, the natural reaction to seeing these new apps is that they are just a 2.0 slightly tweaked version of these things. And the argument there would be, do people really care about accessing Canva inside Chat chat GPT when the tool and experience that Canva has built on canva.com is so much richer and more complex?

10:10Basically, is the convenience of accessing it through chat GPT worth all the things that you lose by losing out on those interface elements? I have a couple thoughts on this. First of all, I would say that while this is a natural first reaction, there are a number of folks who immediately saw that this feels like a different type of thing. Sean Wang or Swix says, okay, two years on, the new ChatGPT apps SDK is much more fully fleshed out with integrations. This isn't Canva, it's Canva inside ChatGPT. This isn't the ChatGPT you grew up with. I'm a little more mixed. I think that it is very likely that there are certain types of apps that work in ChatGPT and others where they either don't work at all, or there are just some very specific use cases that represent a small slice of the overall use cases of the application that makes sense in ChatGPT.

10:59For example, I was not very compelled by the demo they gave where after brainstorming about a dog walking business, they asked Canva to create a logo from within ChatGPT and then turn it into a deck from within ChatGPT. I don't think that there's any universe if you're actually trying to build a business that you're just going to take the few suggestions that you got in ChatGPT instead of going into Canva and using that full set of tools that they have access to to really perfect those things. Maybe as a brainstorming tool without having to switch context, but for final production, there's just no way.

11:28But again, if you go back to what I just said, that doesn't mean that there aren't good use cases for Canva. I just don't necessarily think designing a logo for your business or especially designing a full pitch deck for your business are those use cases. There are probably plenty that I'm not thinking of where the value of not having to context switch between different applications is high enough that people are going to be really excited about that integration right there. What's more, some of the other demos were much more compelling to me. The two others that I mentioned were the interaction with Coursera, where the person who was doing the demo stopped the educational video that they were watching to ask ChatGPT to explain more about what they're saying right now.

12:05And because of the way that they had designed the apps SDK, ChatGPT actually had that context and was able to answer that question. For an educational use case like that, that really leverages ChatGPT's ability to be a personal tutor and assistant in a way that is very different than experiencing that same content over in the Coursera app. The fact that they've designed the apps SDK to talk to apps, again, to use their phrase, so that you're actually interacting with two separate things at the same time, first ChatGPT and second, the application, but also that they can interact with one another is really powerful and I do think opens up some use cases that make apps seem extremely valuable.

12:42I think that the Zillow example is another one where it's smaller, but still, the fact that you can interrogate ChatGPT about all sorts of things around a particular house or property that you're interested in, really is a different experience than you can get in that Zillow app. So whereas with Canva, my perception is that you're making a bunch of trade-offs to experience the app from within ChatGPT, effectively leading to a question of, is the convenience of not visiting the Canva app worth all those trade-offs? In these two cases, for both Coursera and Zillow, there are fundamental things that ChatGPT is adding to the experience that aren't available on those other experiences.

13:17So no, when push comes to shove, I don't think apps are just GPT's 2.0. I think that some will work and some won't work. I don't think it's just going to be an a priori explosion of great apps, but I can see how they become a much more can't believe they weren't there the whole time type of part of the ChatGPT experience. Overall, it feels to me like there was some amount of a vibe of this being a shift from big innovation to more practical integration. Dan Shipper from Every wrote, OpenAI launched a lot of exciting stuff, but it feels less exciting for developers and more for developer-adjacent roles in the org.

13:51You should be hyped if you're doing AI ops in a company, but if you're a hardcore AI engineer, it's a bit underwhelming. During the Codex section, he followed that up with saying, Codex CLI is my daily driver, but to be honest, the updates seem pretty incremental. And in fact, combining this with the next question of what's going to have the biggest impact, a lot of the developers who were there were most stoked simply about the updates to the API. Writes Matt Schumer, both GPT-5 Pro and Sora 2 are coming to the OpenAI API today. These models are both massively better than what developers had access to just a day ago.

14:23We're going to see some very interesting effects from this. Now, he did also point out that the price is significant with GPT-5 Pro 12 times the price of regular GPT-5, but at the same time for some use cases where GPT-5 Pro is so much better that those use cases come online for the first time, that could be worth it. Matt also noted that in addition to Sora 2 coming to the API, Sora 2 Pro is coming to the API. Now, we did not get confirmation of this, but one of my theses after having used Sora 2 for a week or whatever it's been, a handful of days, I have no sense of time anymore. It's felt to me very much like there was a Sora 2 Pro out there that we were not getting access to in the app.

15:02And so I'm excited to see if that is actually the case and how much better it might be. So coming back to this question of biggest impact and this idea that maybe we're moving into an integration phase, I do think that while a little bit this is a false dichotomy, there is definitely a more practical bent to the things that companies are releasing right now because there has to be. There is so much real usage happening that people aren't just blindly impressed by every new parlor trick. They are really working hard to get work done or to just do whatever they're trying to do with these tools. And so these updates that seem incremental are actually just pretty essential for making these tools live up to their promise.

15:38It may not be as big and splashy, but it is the phase that we're in. And it's the stuff that's going to unlock a lot of the actual real lived value of these things that goes beyond just the demo stage. So for my money, when it comes to the potentially biggest impact of this, as much as the agent builder, I do think is a big deal and will help start to see a proliferation of agents. I kind of just think that that's a natural waypoint along the road. I think OpenAI was always going to be deep in the agent infrastructure and building game. And I don't think that AgentKit, as it's expressed and shared currently, is going to all of a sudden overnight unlock hundreds of thousands of new consumer agents coming to market.

16:15What I think is interesting from a competitive landscape perspective is the potential that apps turn into a context black hole where OpenAI simply sucks in all of the context and information and creates just an absolutely enormous advantage for themselves when it comes to consumer lock-in around ChatGPT. Let's once again hold aside the example of Canva, where you're making a trade-off between the convenience of just doing it in ChatGPT versus the full feature set that you get in the Canva app. And let's instead look at the Coursera and the Zillow example. It strikes me as quite possible that once you have used Coursera, with your personal ChatGPT tutor assistant sitting next to you, you will not want to use Coursera in the normal form.

17:01Now, certainly that doesn't mean that ChatGPT couldn't just go live in Coursera as well, but in either case, it locks in ChatGPT as the personal companion experience when you are using that app. Likewise, with Zillow, I don't think that we're going to see people abandon their wistful browsing. And to the extent that they're just using Zillow to get housepiration for their next few hours of work, I don't think you necessarily need the type of research feature that you get with that ChatGPT integration. But for someone who is actively looking at properties, trying to understand where they sit, how they're going to meet their needs from a life or a business perspective, once again, it seems hard to go back to the world where you're just doing it on your own versus the world where your personal assistant is sitting there with all the context that it needs to actually help you find out what you want to find out and make better decisions.

17:48And so it strikes me that apps might be a sneaky way to get closer to ChatGPT's vision of a real true assistant for every person. Like with everything else, I don't think it's a foregone conclusion. I think we have a lot to learn about which types of current application experiences really benefit from that side long assistant. I could be being seduced because it's so clearly valuable inside the context of education, but I think it's something really interesting to watch. So anyways, friends, those are my first thoughts, my first reactions to OpenAI Dev Day. Let me know what you think. And as I mentioned, we will be back tomorrow with a full deep dive, a recap of additional announcements from throughout the day, a look at the big AMD deal that OpenAI did, and how the conversation has shifted with the benefit of a few hours of actually digging into all these announcements.

18:33That's going to do it for this special bonus episode. Until next time, peace.