MCP, Agents and What AI Engineers Are Thinking About Right Now feat. Swyx

0:00Today on the AI Daily Brief, what non-engineers need to know about the state of the discourse in AI engineering. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. To join the conversation, follow the Discord link in our show notes.

0:17One of the things that I think is very exciting about AI is that it's breaking down barriers between technical and non-technical people. AI is an intermediating technology whereby people who are non-technical can start to grok the tools of creation from an engineering and development perspective. And as people are trying to make that side of their brain work, the resource that I most often point them to is the Latent Space podcast and newsletter and the AI Engineer Summit that's produced by some of the same people. Specifically, today's guest, Swix, is at the center of all of that amazing work.

0:51He has an incredibly good pulse on the state of conversations when it comes to AI engineering. And so today we're talking about what the big themes that people in that community are talking about, building around, and what the implications are for the rest of us. All right, Swix, welcome back to the AI Daily Brief. How are you doing, sir? Very good. Long-time listener and glad to be back. Yeah, so I think that this would be a really fun conversation. What I was saying to you kind of before we were recording is that I think that what's super valuable and what I hope to kind of have come out of this is to help my listeners, which I would say are on average, there's a higher portion of non-engineers than your audience.

1:31So, you know, helping enterprises and non-engineers understand kind of where their big discussions in AI engineering are. And I think that it's pretty clear at this point that, you know, it has always been valuable for people who are inside technology and building with technology, whether they're engineers, whether they're developers or not, to try to keep a pulse on what builders are building, how they're building, you know, what tools are using. I think it's even more pertinent, obviously, with AI, right? That like the space between the non-engineer and the engineer is getting blurrier, right?

2:02perhaps to some chagrin somewhere. But so I think that that would be super valuable. And I think you obviously have a unique vantage on this, you know, in terms of content you produce, you know, with latent space, but also through planning the AI engineer summits, right? So we just got off one, I guess a couple months ago now, it feels like just a minute ago. That was super fun in New York. You've got the AI engineer World's Fair coming up in San Francisco again this summer. And so I thought what would be fun is maybe we kind of just go through, like use those planning processes to kind of frame what people are thinking about and how that's changing even in this compressed period of time.

2:42And maybe to start off, I think what you were just sharing with me about how you think about planning, I think is actually very useful context for folks to get into the conversation. Yeah, sure. Thanks. The planning process, this is my third year doing this, So I don't feel like I have it fully on lock. But, you know, I think the main thing is that our source of alpha is that we stay close to the engineers. And also we react faster than the machine learning conferences. And so these are the two things because there are competitors. There are obviously many, many conferences. conferences, but the, for example, the research conferences like NeurIPS, ICML, ICLR, all of them, I don't know if people know, in order to speak at one of them, you have to submit a paper six months in advance.

3:32And so in AI time, that's a long time. And that's just purely because NeurIPS is 38, 39 years old. It just wasn't that fast when they were started. And now it is, and, you know, it's hard to change a tradition like that. And then the other conferences are typically organized for business heads and talking heads and people interested in that kind of thing. And so they don't get too far down on the technical detail. And I think the key problem with that is there's a ton of fireside chats, a ton of panels. Everyone shows up with no preparation whatsoever. They yap for 30 minutes and then you're done and you don't remember any of it.

4:09So the thing that I emphasize a lot is what are engineers going to take home with them to do their work and how to improve that. And that means I demand of my speakers that they prep a lot. But then also that gets the results that we get, which is talks that actually matter. And the people that come actually want to meet the folks that build things and are hands-on. So yeah, I mean, it's weird because it's slightly less prestigious to be hands-on keyboard than being a CEO of a major company you know uh going on stage and talking about how we're all not going to need jobs in five years but uh you know the people who need who who are hands on keyboard also need a place to gather and that's what i do so yeah and my argument as someone who's you know uh been involved very nominally you know at least with the last one you know helping mc and stuff is um i think that in general the action is happening hands on keyboard you know and uh and even if you are you are everyone's a builder now is sort of the short of it, you know, whether, whether they're building with code or building other ways, if, if you are kind of fully participating in AI and agent land.

5:21So let's talk about going into the, the sort of the, the, the summit that was earlier this year. It felt like the big sort of inflection or change again from the outside. And you correct me if, if you were thinking about it differently is an expansion of the question that had kind of characterized a lot of 2024, which is what is an AI engineer, right? And what does it mean to do that? And what do you need to think about? To what is an agent engineer? And how does agent engineering interact with, change, modify, transform that framework? So I'd love to know if that was sort of how you were thinking about it, if that was the big sort of shift, and what the implications were in terms of the conversations that you wanted to facilitate.

6:05Right. Yeah. I don't know if people understand that it felt like a risk at the time because we made this decision kind of November-ish. And for a lot of last year, actually, agents was kind of a bad word because there had been a few agent startups that failed. And people were like kind of taking it. We were telling people to take agents off of their description because it wasn't so ill-defined and so overused that people didn't really like seeing them anymore. It was a counter signal that you were doing something interesting. And then it really flipped with O1 and with all the other subsequent agent launches with operator deep research and all that.

6:49And Manus now is crazy. So like last year's World's Fair, we had nine tracks. Only one of the nine tracks was agents. So we really had to decide, okay, this is the right time for agents, and we're going to go all in on this one. And I think then there's also another consideration with regards to what can engineers uniquely do as opposed to researchers. And we had other talks about open models. We had other talks about GPUs and inference and multimodal models and all that. but a lot of that starts to entangle with the sort of research layer of the stack and those you know those like those are those are great but like they are very dominated by people with the resources to do that research and there are already research conferences so we really wanted to be an engineer conferences and I think that specialization in the engineering layer on top of models to turn them into agents was like the key like I was like I was kind of waiting for that moment and it felt like the time to do it, especially with NCP's launch, like, end of last year.

7:55And so I picked, you know, I announced that we were all in on agents and we, like, planned out, you know, here's, like, what we think the disciplines of agent engineering are. It turned out to be very different in the end, but, you know, we sort of scoped out what that call for proposals was. And people came in and it was really popular. So, I mean, I can talk through, like, the individual talks that did well, but also, like, That's the high level, which is we made a high-level bet on agents, and then I did my keynote with why now. I think there's a very strong moment to timing where if you're correct but too early, you're still wrong.

8:34And I think this whole trend of 2025 being the year of agents, I think it's probably correct. what do you think that you know without rehashing the entirety of the of the keynote what do you think the sort of the key inflection point was was it the like the reasoning models was it you know better infrastructure i mean what what do you think sort of made that that shift for it to become real uh if i can share my screen i'll um i'll just you know for people on on youtube um i had i had sort of nine points uh and a lot of them are more slow bake right like these these These have just been improving in terms of performance and the model of performance.

9:12So what I'm looking at on the screen now is the Gaia benchmark for meta. By the way, we interviewed with the Gaia team last year if you want to learn more about what Gaia really is and what their intentions are. But things have just been improving on this S-curve. And we're just kind of in this top part of the S-curve now where it's starting to reach human baselines. And I think the closer we are to human baselines, the more we can start actually using them instead of just reporting their benchmark scores and saying, that's cute, but I'm going to go back to using my human intuition now. But then there's also all these other stuff, right?

9:44Like there's better capabilities, better tools, better tools. But also like, you know, I like to emphasize the second tier of stuff because I think people aren't really, people always focus on the first tier, which is that, oh, we've got, you know, a reasoning model now versus non-reasoning models. And that makes all the difference. And like, sure, like that helps, but you could build agents without reasoning models and still benefit from all these other things. So like multiple frontier labs, like Grok 3 now is an API. You know, Gemini 2.5 Pro is arguably the best model in the world. Like you're not stuck to one model, therefore you can chain together different capabilities and get out of ruts.

10:23The depreciation curve of models is also a constant force, right? And it's all Moore's law and all that other good stuff that we can talk about later. And I think the last thing on the business side I want to highlight to folks is that we're actually really moving from a cost plus model where you're just charging based on a number of tokens and then maybe you mark it up a little bit towards the outcomes that you deliver. And that's a huge change, right? Because now the reframing is going from, all right, how well can you consume my tokens to how much of my job as a human can you do? Therefore, you are worth this much.

10:59And that's a couple of orders of magnitude difference there. Yeah, super interesting. And so kind of what were the types of talks that you were trying to bring together to, you know, instantiate this and bring it to life? And what hit either, you know, and particularly I love this sort of subjective take on, you know, what was popular that you didn't expect or did expect, you know, out of the talks that kind of were most resonant with people? Yeah. Well, I mean, so we're still in the process of releasing all the talks. So I don't know in advance with everything. but you could see like anything from Big Lab is good.

11:36Sure. We had a, because we were organizing in New York, we really wanted to also focus on agents in production, right? I think the subtitle of the conference was agents at work. So really, I think a lot of people see demos and then they're like, that's cute. I can use it for fun demos and then probably they never actually use it. But who's actually using this thing at work? How much impact is it having? Do any of the Fortune 500s care? you know, like what have they figured out because they're so smart and so big and so much, so resourced, what have they figured out that I haven't figured out, right?

12:10So I got people from like Jane Street, from Bloomberg, from BlackRock, which by the way, their talk wasn't approved to be released. So everyone who attended got an exclusive that we cannot release. Alphabet. And Ramp as well, yeah. So the production agents and AI talk in big companies like Jane Street, Bloomberg, Ramp. Ramp, by the way, I think announced an$11-12 billion valuation after this talk. No correlation there. And then I guess RL was very hyped, and we can talk about that one, but also Windsurf, which I think always very surprising to me that you can just be a second mover after Cursor and still do super well.

12:59and and you know as long as you you design your agents well and you're you you take like good enough uh differentiation you have good enough differentiation people will give you a shot and i think that's very encouraging i think that that just means that you know if you think something is over or a category is done maybe you should just try harder yeah uh yeah what when i i was sort of pulling out uh and looking at you know again sample that not everything has gone up but some of the standouts included the RL for Agents, the Windsurf thing, and then, of course, and maybe the one that we can talk about a little bit now is the MCP discussion.

13:37So that one I expected I expected it to do well, so it is not Yeah. Well, so I would love to, one of the things that was really interesting is you wrote a post, or you guys wrote a post called Why MCP Won, which I think is super interesting. I think I did a whole episode about it, basically, that specific post. Yeah. I mean, listen, you keep creating great content. I will keep, you know, remixing it for this audience. So I thought part of what was interesting because it's so, so recent, but still so far in my memory now that MCP has completely taken over the conversation, right? And you had the Google CEO a couple of weeks ago asking to MCP or not to MCP.

14:15And then yesterday answering that question or, you know, yesterday from when this was recorded. But you were you were reflecting on the sort of, you know, the the the period following the immediate reaction to it. So you basically argued that the immediate reaction was good, but then it kind of, you know, got quiet for a little while. So can you take us back to like, you know, end of November when it gets announced into December and January and how you were thinking about MCP and then where you saw the sort of the pickup in conversation and what you attribute it to? Yeah, I would give credit to Alex Albert, who I think has a timeline of events of MCP somewhere out there.

14:50But yeah, it was launched in November. And I think there was a lot of interest. I think it was top of Hacker News. But I don't think there was a ton of immediate follow through because people were kind of used to big companies launching protocols and then it's kind of flopping or not really working. One recent one that people might not remember is Meta actually launched a Lama stack, which is a full open source framework and stack. And every framework has a protocol embedded in it. So I mean, that was kind of the insight there, that everyone maybe went too far in trying to impose all these opinions at once.

15:31And Anthropic took a different approach and adopted a protocol that other frameworks could build on top of. And that maybe was that minimal viable product that actually was the only viable product because everything else would have been too much for imposing too many opinions on everyone building stuff. So I think a lot of people started exploring it and I think integrating it into their workflows. And I think probably it was driven by the IDEs. So like Zed and Winserve and then eventually Cursor, I think, was the last one to add it. Maybe Copilot as well. I'm not really sure the exact sequence there.

16:06But I think the, yeah, I mean, really, I knew that it was going to be interesting. I thought it was useful for a big lab to come out and announce something that wasn't a model and it was how tools should interoperate with models. And I think insofar as OpenAI did that in 2023, 2024, with their function calling spec and tool calling, which did well but didn't have that spark of excitement that MCP had. I think Anthropic taking a stab at it was really strong and I wanted to feature them. And that was really about the amount of calculation that I had about there. They really took it all the way. Anthropic has been a really strong supporter of my conferences.

16:55so they showed up and they had this like two-hour presentation and they had tons of new alpha they never dropped anywhere else and then they also talked about their future plans announced the official mcp registry at the conference and it was all this stuff and so like that sparks in more excitement because i think the other thing about launching these things from big labs is that they need follow through they need to people need to believe like this is an actively worked on thing you know i think one of my statements that that you liked in the in the in the piece was that you know protocols are only as strong as the people who are already using them um and so you just have to believe that if i invest in mcp all my buddies are going to invest in mcp all the people that i want to be compared to are all investing in mcp um and so like yeah i mean you know that that workshop did did really well we released it it was uh it was like a good hit and i think like i saw the numbers earlier than anyone else just because i could saw i could see like the views i could see the i looked at the download statistics and i looked at you know everything where everything was trending um so i think this was the chart um that um i focused on and i was like okay you know do i call it now is it too early to call it it was it was like this is exactly where you know it's like three four months into mcp there have been many many attempts at creating some kind of agent benchmark uh agent standard but um you know nothing like this and i was just like oh you know i think it's a decent chance that um mcp has kind of won this and i would try to articulate to myself of why it won.

18:24And I ended up with these seven reasons, right? Six reasons, which is like, it's AI native, it's open standard, blah, blah. You already went through this in your podcast. And guess where, I don't know if you've seen the charts since this post. I have not. Yeah, well, we can click on it. And it has done, we might need, it might need some loading time for the data. But basically I projected that MCP would take over the incumbent of open, open 8, so this is just GitHub stars, right? So we are going from 0 to 15k in a very short amount of time and faster than anyone else. But the incumbent is OpenAPI.

19:02That's the big behemoth that is basically the old industry standard. And that one is at 30 ,000 stars. So it's basically continued to go there. So I was doing a conservative projection and I was like, it'll hit 30 ,000 stars in July-ish. No, it's crossed it this month. It's crazy. Today's episode is brought to you by Vanta. Vanta is a trust management platform that helps businesses automate security and compliance, enabling them to demonstrate strong security practices and scale. In today's business landscape, businesses can't just claim security, they have to prove it. Achieving compliance with a framework like SOC 2, ISO 27001, HIPAA, GDPR, and more, is how businesses can demonstrate strong security practices.

19:45And we see how much this matters every time we connect enterprises with agent services providers at Superintelligent. Many of these compliance frameworks are simply not negotiable for enterprises. The problem is that navigating security and compliance is time-consuming and complicated. It can take months of work and use up valuable time and resources. Vanta makes it easy and faster by automating compliance across 35-plus frameworks. It gets you audit-ready in weeks instead of months and saves you up to 85 % of associated costs. In fact, a recent IDC white paper found that Vanta customers achieve$535 ,000 per year in benefits, and the platform pays for itself in just three months.

20:21The proof is in the numbers. More than 10 ,000 global companies trust Vanta, including Atlassian, Quora, and more. For a limited time, listeners get$1 ,000 off at vanta.com slash nlw. That's v-a-n-t-a dot com slash nlw for$1 ,000 off. Hey, listeners, are you tasked with the safe deployment and use of trustworthy AI? KPMG has a first-of-its-kind AI risk and controls guide, which provides a structured approach for organizations to begin identifying AI risks and design controls to mitigate threats. What makes KPMG's AI Risks and Controls Guide different is that it outlines practical control considerations to help businesses manage risks and accelerate value.

21:02To learn more, go to www.kpmg.us slash AI guide. That's www.kpmg.us slash AI guide. Today's episode is brought to you by Superintelligent and more specifically Super's Agent Readiness Audits. If you've been listening for a while, you have probably heard me talk about this, but basically the idea of the Agent Readiness Audit is that this is a system that we've created to help you benchmark and map opportunities in your organizations where agents could specifically help you solve your problems, create new opportunities in a way that again is completely customized to you. When you do one of these audits, what you're going to do is a voice based agent interview where we work with some number of your leadership and employees to map what's going on inside the organization and to figure out where you are in your agent journey.

21:53That's going to produce an agent readiness score that comes with a deep set of explanations, strength, weaknesses, key findings, and of course, a set of very specific recommendations that then we have the ability to help you go find the right partners to actually fulfill. So if you are looking for a way to jumpstart your agent strategy, send us an email at agent at bsuper.ai and let's get you plugged into the agentic era. Yeah, I mean, it was to me, you know, picking up on the signals strictly from competitive pressure when, you know, because OpenAI was fascinating because when they announced agents SDK, it was sort of like, all right, cool.

22:33There's, you know, the natural interpretation, I think, you know, perhaps unsophisticated was, got it, there's going to be an agent sort of protocol war, right? Like, that's another vector of competition in terms of, you know, developer allegiance that we're going to go for. And then when like five minutes later, they were like, we love MCP, we're supporting MCP too. I was like, all right, well, that's a whole different Paul game. Obviously, Google then follows. And, you know, I did resonate with that point that you made in terms of basically the network effect of protocols. Like, there is such huge advantage, obviously, for building where other people are building.

23:07And it just got over the bootstrap problem so quickly that, you know, it was almost like this is the type of thing where if everyone can get together more quickly and make the decision, it's so good for everyone in terms of just collective value that everyone provides each other, you know. And now you've got people who are actually thinking about, this is a category of new startups, not just a new tool to use, but an actual category of things to build. You've got Dharmesh Shah, who I know is on the show recently, too. He's spitting out on LinkedIn all the MCP-related startup ideas that he doesn't have time to do.

23:41And I think he funded one of them when someone responded saying that they were doing that thing. Oh, wow. Awesome. Yeah, so it's very cool to see how fast that ecosystem has emerged and is starting to flourish. Yeah. So I think people think of me as an MCP show just because obviously I did just show it a bit. But I am a bit measured about it, right? Like I have seen protocols get hyped and then get very not hyped. And, you know, so the most recent version of this in developer land is GraphQL. There was like, yeah, this is like a better layer over REST. everyone's doing like rest versus GraphQL threads and all that and it's very reminiscent of MCP versus OpenAPI it's like basically the same and all the by the way all the issues that came up with GraphQL also are emerging with MCP like how do you do authorization how do you connect with remote MCPs and do discovery of them like the exact same things because these are all the same type of problems which I call the m times n to m plus n problem right like that's how you sort of You solve combinatorial issues by adding one layer of abstraction that has a standard interface that everyone plugs into, right?

24:50Like common concept, everyone understands that. The authors of MCPR are super aware of it. They talked to me about it after we did a podcast with them. So, I mean, I think good governance and good judgment is still going to win the day. There is a way that they can screw this up. and I think I was also on these TBPN podcasts and they were like, is this going to result in like an explosion of agents? And I'm like, there's already an explosion of agents. This is not really changing that trajectory in any way, but this basically improves the quality of integration. So it's a boring answer. It's just like, you know, the integrations that you write, or you expect in one app, you're going to see in another app because it's super easy to add them.

25:30And you don't have to wait for them to add like Notion just because it's on their backlog and they don't have it prioritized yet. Like, no, it's just out of the box because Notion just launched an MCP yesterday. And that's it. Like, it doesn't make for super agents or anything apart from they are wider, they can integrate with more things, but we still have to solve a lot of the other core problems of agents. So I think it's a good caveating. Where I look at it from is, so, you know, again, bringing this sort of back to an enterprise audience, a lot of what I think enterprises are trying to figure out right now is, interestingly, when they think about agents and agent capabilities, I actually think that directionally they're correct.

26:12When they're imagining in their kind of brain and ideal mind's eye of what agents can do, they're kind of right about where things are headed. The problem is just that it's not there yet in most cases, right? The things that agents can do are more limited. They're more discreet, you know, yada, yada, yada, right? There's some gap between what they're imagining and really excited about and what they are now. and um a big calculation is how much how fast and in what way to invest given the the rate of change and this is really really challenging because it is sort of obvious that the answer is uh it can't be for in most cases just practically go all in on building the thing that you most want to build right now because in many cases it's just not exactly where where where they want it to be but it also can't be on the other end of the spectrum just wait for it to get ready because you're going to be behind by then.

27:00And so I think that they're trying to understand what to do in the interim. And so it's really interesting when you have, call it accelerationist forces, which is sort of another way of describing, I think, what you just described with MCP. It's boring because it's not going to make more agents, it's not going to change the trajectory. But by having, to your point, not having to wait around for that notion of integration, not having to wait around for some other thing that you're waiting for, it does feel like it is likely to accelerate the speed at which new capabilities come online. And each new thing that gets connected to the ecosystem is likely to open up some additional use case.

27:37Yep, I broadly agree with that. Yeah, we can talk about the other elements of agent engineering that we got from the talk, from the conference. Yeah, I think MCP is a great protocol, but imagine if there were standards for all the other stuff. That would be great. Let's talk about the outside of that, the other pieces. Yeah. Well, so the other thing that happened at the conference that was followed up after was OpenAI actually previewed how they think about agents, which is they released – this is the OpenAI for VPs of AI talk, the one that was the day before you came. and they said an agent is an ai application consisting of one a model equipped with two instructions that guide his behavior three access to tools that extend his capabilities that's mcp four and encapsulated in a runtime with the dynamic life cycle um so that was what they previewed and then they launched the agents sdk after that they i mean they told me that that's what they were doing so um yeah i i had i had full knowledge of that so um it's interesting that That is one form of the definition.