Skip to content

By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Product SchoolProduct School

Zapier CEO on Rebuilding Products to Be Agent-First | Wade Foster

Every time OpenAI or Anthropic ships a new model, a wave of startups dies. Wade Foster keeps finding ways to stay alive.

Carlos González de VillaumbrosiahostWade Fosterguest
Nov 18, 202632mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

Great conversations run long. Their best ideas don't have to.

uListen is free on iPhone, iPad, Android, Chrome, and Claude.

See how

Transcript

  1. 0:00 – 0:53

    Trailer

    1. CV

      Every single time I see a news of OpenAI or Anthropic release a new variation of their LLMs, a bunch of startups die, but somehow you keep finding ways to, to stay alive.

    2. WF

      It's like having the world's best developer, like, sitting right next to you and you just tell him, "Hey, can you build an automation for this? Can you connect this to that?" Turns out the best no-code product was always just an agent writing code for you.

    3. CV

      Wade Foster, CEO of Zapier. Why can't OpenAI or Anthropic crack that code yet yours seems to be the winner?

    4. WF

      Fortune 500 company that went from saying like, "Ah, we're kind of nervous about this," to now they say, "Hey, we do not let any connector or automation run on any system other than Zapier."

    5. CV

      Today, a lot of these enterprises are still using AI assistants for use cases, but the actual connection in between other tools to automate work is still something that is not happening mostly because of fear.

    6. WF

      I do empathize with the leaders in this position. You really feel like you have one of two op- One is to-

  2. 0:53 – 2:44

    Intro: Carlos welcomes Wade Foster, CEO of Zapier

    1. CV

      Hey, this is Carlos, CEO at Product School and your host on the Product Podcast. My guest today is Wade Foster, CEO of Zapier. I've been using his product since the early days and he's one of those rare founders who keeps surviving every model release that supposedly kills his category. Fully remote, profitable, and here's the twist. Zapier grew up as the no-code company and it is now a code product. The users just never see it. Here are the things we'll cover. The GPT-4 launch was the real wake-up call, not ChatGPT. The best product was always an agent writing code for you. Don't build an agent. Have the agent build a workflow. About 71% cheaper. How governance turned the scariest thing about his category into the moat. No meetings before 11am and why he rarely does one-on-ones anymore. Let's get into it. If you're a product manager at a private tech company, stock options may be part of your compensation. Exercising those options lets you own shares in the business you're helping build. But becoming a shareholder can require a substantial amount of cash. If exercising costs $50,000, that's 50,000 of your own savings committed to one company before any related taxes. Our sponsor, EquityB, provides employees and executives holding stock options with the capital to cover their exercise costs and related taxes without using their savings. This is non-recourse funding with repayment tied to a successful exit. So if the company never exits, you owe nothing. EquityB has already helped thousands of employees at more than 900 companies access their equity. That gives you a way to turn your options into shares while keeping your savings available for other investments or personal goals. You still have a stake in the company you're helping build. Explore your funding options at equityb.com/ps. That's equityb.com/ps.

  3. 2:44 – 3:38

    Welcome, and why Zapier keeps staying alive

    1. CV

      Welcome to the Product Podcast, Wade.

    2. WF

      Awesome. Thanks for having me, Carlos.

    3. CV

      Man, I've been using your product since the early days and I feel like you are one of those gladiators that finds a way to stay alive. 'Cause every single time I see a news of like OpenAI or Anthropic released a new variation of their LLMs, a bunch of startups die, but somehow you keep finding ways to, to stay alive.

    4. WF

      Yeah, I mean it definitely feels like at least the algorithms right now it's, you know, we kind of seesaw back and forth from the, uh, it's so over, we're so back [laughs] mentality. So yeah, a lot of that, you know, just feels like kinda like the algorithms and the headlines kinda yanking things around a little bit. Um, our approach has always just been wake up every day and just, you know, listen to our customers and figure out how to ship things that make their lives better. And I think if you do that, like, over the course of decades, like, your, your chances are pretty decent that you're gonna, uh, still find a way to be in business.

  4. 3:38 – 3:48

    Fully remote and profitable

    1. CV

      And I think there are a few unique facts about your business. First of all, you guys are fully remote, right?

    2. WF

      We are.

    3. CV

      And you are profitable, which is a rare thing to, to see these days.

    4. WF

      We are profitable

  5. 3:48 – 4:33

    The repositioning: baking AI into the core product

    1. WF

      as well too.

    2. CV

      Tell me more about that name reposition or expansion if you will, that you had to make in your product to make sure that you are incorporating all of these new AI capabilities.

    3. WF

      Zapier, you know, we grew up in kind of the, the no-code era. Uh, you know, a lot of folks kind of credit us with being one of the pioneers for that stuff. And mostly the approach was how do we take a lot of the burden of creating automations off of humans who don't know how to write code? Like, it was like let's g- let's make it easy for them. And so for the longest time, that was putting, like, these UI abstractions on top of APIs, you know, trying to make these simple point-and-click tools that took away,

  6. 4:33 – 5:55

    Why GPT-4, not ChatGPT, was the real wakeup call

    1. WF

      you know, the, the, the toil that someone who didn't know how to write code had to do to sort of build these automations, build integrations, et cetera. You fast-forward to 2022, 2023, G- ChatGPT launches and, yeah, I, I, wouldn't say that that was, like, the sort of r- first, like, true wake-up call. We were like, "Wow, this is a really impressive product for sure." But for us, the power really dawned on us at the GPT-4 launch, which happened like six months later. GPT-4 was a much better model than ChatGPT launched with. It was a lot cheaper. And the fact that it launched only six months later felt to us a sign that there is about to be this like, you know, exponential curve of improvements that happen with these models. And so we thought, "Hmm, well if that's the case, what needs to change about Zapier?" And it felt like there was two really important dimensions to this. First was the product needs to change. Like, we need to bake these capabilities into the core experience. The second thing was how we work has to change. Like, these tools can help us be more productive, work better, et cetera. And you know, I'd say i- initially this was like a period of quite a bit of exploration inside of Zapier where, you know, we were trying a whole bunch of new capabilities, shipping things inside the product, shipping how we worked at the company. But by and large, I, you know, I'm trying

  7. 5:55 – 6:40

    The turning point: models one-shotting code nobody reads

    1. WF

      to think like how many of those experiments actually survived. Um, and I, I'd say most of those It actually didn't. You know, it really felt like the turning point happened about a year ago when, you know, uh, the Opus 4.5, uh, GPT-5.2, I think or whatever came out, 5, 5.0 came out. I, I forget the names of them. [chuckles] I c- it's hard to keep track. But the, you know, these, these were, like, really sophisticated reasoning models, and we saw all of a sudden that you, you know, at least internally, like, these things were one-shotting, like writing code. And we were like, "Okay." Like, it's, it's really here. Like, you don't have... Engineers were not reading the code anymore. They were just, you know, prompting the model. The model would write it. They would have some

  8. 6:40 – 7:43

    Why Zapier is now a code product users never see

    1. WF

      tests or evals or, uh, ways to sort of check its work, but by and large, not reading the code to do that. And that, to us, was finally the first moment where we're like, "Ah, we know how the product can change from here." You know, where it once was no code was about abstracting away the toil of working with code, we realized, okay, like Zapier is actually going to write the code for you. And so now when you look at what is growing the fastest inside of Zapier, it's, um, it's Zapier MCP. And the reason it's growing so fast is you can install it into your agent of choice, and then the agent goes and builds the automations for you, and it writes it with code. So Zapier is a code-based product now, even though the users aren't. It's like having the world's best developer, like, sitting right next to you, and you just tell him, "Hey, can you build an automation for this? Can you connect this to that? Can you do these things for me?" And all of a sudden, it, it sort of brings to life what a lot of the original vision of Zapier was, which was make automation work for everyone. And it turns out the best no-code product was always just an agent writing code for you in the first place.

  9. 7:43 – 10:09

    Why OpenAI and Anthropic can’t crack it: focus beats general

    1. WF

      [chuckles]

    2. CV

      It's so funny because you're quite right. Your product is more technical than ever, yet your audience, your main customer is still the non-technical user. And so my question to you is, okay, you make this MCP work, and now people can use your product from other interfaces. Why can't OpenAI or Anthropic crack that code? Because they keep releasing their own vers- variation of their automation workflows, yet yours seems to be the winner.

    3. WF

      I think there's kind of like a product truism here, which is that a focused team on a very specific set of problems is always gonna be a little bit ahead of a team that's, like, working on more general capabilities. Yeah, I think OpenAI, Anthropic, they're building, like, these incredible models and these incredible capabilities. They've got even some amazing applications on top of them. But they also build a lot of stuff that doesn't work either on the application side, and they're quite large companies now, too. And so, you know, the things that they really care about are gonna be excellent, AKA the models. The models are just continue to just get better and better and better. But I still believe that there is a huge amount of room for entrepreneurs to build on top of those capabilities and extend what they can do into specialized dome- domains, into unique use cases, um, build novel UIs or, you know, user experiences alongside it that won't come from the labs. And so that to me is where it, it is definitely, I think, for every entrepreneur I talk to, like, a, a, a somewhat scary time because you sort of think, like, "Hey, are they just gonna eat everything we do?" But again, I think if you kind of come back and pay attention to, like, where they're innovating and say, "Hey, how do we go alongside of them instead of in front of them?" You know, you're gonna find yourself in a situation where you can build capabilities that continue to be complementary over time. I wish I had, like, a silver bullet to sort of propose to folks and say, like, "Ah, if, if you do this, like, you will, you will always have a complementary product." But y- you know, what I see, like, the best companies do is they're just very quick at, like, rewriting their priors. When new model capabilities come out, they're, like, willing to just say, "Okay, throw this experiment away. This doesn't matter anymore. Let's shift over to this thing because this is where it's all gonna, gonna work." And if you can build a community of customers that sort of appreciate, like, your orientation, you'll always be at least a step ahead on that, on that specific dimension. You may not be on, you know, this other set of things, but the thing that this group of customers cares to have the best thing in the world for, you, you can really make sure to do a good job at that.

  10. 10:09 – 11:09

    The two strategies: build against the models or alongside them

    1. CV

      Reminding me of something that Sam Altman said long time ago, which he said there are two kinds of AI strategies. One which is w- trying to build your product kind of against the models or betting that you can have an edge and do something better than the models, or the other one is to-

    2. WF

      Mm-hmm

    3. CV

      ... assume that the models are going to get so much better that you can, y- you have to work alongside them, right? And I think that's similar to what you are saying here, is assuming that the models keep getting better and then picking where you can really add the value.

    4. WF

      Yeah. I think that's what it really does boil down to, is that you, you have to p- pick, like, where you think that there are places that you can sort of extend on top of them. And I, you know, I think this does, like, requiring you to, like, really understand how these models advance. You gotta understand the bitter lesson really well. You gotta understand, you know... The, the common trope is, like, how do you build product capabilities that get better with new model releases, not get wiped out by new model capabilities? So, like, that's the sort of, like, fork that you're trying to, like, thread here, uh, when making these decisions.

  11. 11:09 – 12:50

    What being agent-first actually means

    1. CV

      Other thing I want to explore with you is this concept of being agent-first. As more and more agents use the web and use the tools, in fact, they are now more a- there are more agents out there than humans using those products, right? So how are you ad- a- adjusting your product to make sure that this kind of new user is making the most out of your product?

    2. WF

      I mean, I think we're starting to see that agents are really truly gonna be the new computing platform. You know, in the last couple months we've seen products like GroqBots and Muse come out and gain, like, a huge amount of popularity. Uh, OpenAI and Anthropic launched their versions of it, Dots and, um, Duet as well. And so, you know, these are products where humans are basically just delegating their computing work to these agents, and the agents just go and act on their behalf. And to me, that feels like it opens up, like, a new It, it's just sort of like a, it's a new computing paradigm. Uh, you know, in the past it was, "Hey, we've got an operating system, and let's build software for the operating system." And then it was like you had browsers, and let's build, you know, uh, capabilities for browsers. And, you know, then you had mobile, and it's like, "Ah, let's build app for mobiles." So to me it's like j- yet another an extension of that. It's like we've got agents. Like, well, what are the tools that agents need available to them to go com- uh, do well? And so, you know, a, a really interesting way to think about, you know, your opportunity is, hey, here, here is what these agents can do today. Here's what they kind of stink at. Do you think the models or the harnesses are going to eventually build those capabilities? If no, then there's a really interesting product idea, uh, there, where you can say, "I'm gonna build a product that the agent can reach for." And the agent should be able to sign up for it, should be able to buy it, should be able to do everything that a human should be able to do using that product. And if you do that, yeah, I think you've built a product that is effectively

  12. 12:50 – 13:11

    Keeping cost down: model-agnostic and tool-agnostic decisions

    1. WF

      agent-first.

    2. CV

      So as you build your own agent-first product, and you try to be as agnostic in terms of tools as well as models as possible, like how do you go about making those cost decisions to make sure that, you know, guy... people are g- making the, the best possible decision and they are not getting slapped in the face with a huge

  13. 13:11 – 14:00

    Don’t build an agent, have the agent build a workflow (71% cheaper)

    1. CV

      bill at the end?

    2. WF

      Well, uh, that is tricky right now because there is definitely a, a lot of like expensive usage, uh, going on. Yeah, I think one of the things that we encourage all of our customers to do is that a lot of times they're building agents when they should just be b- having the agent build a workflow. We see this ourselves, so like if you... I- instead of asking Dropbot or instead asking Muse or instead of asking whatever your agent tool is to sort of go build an automation for, build a routine for you, have it go build a Zap for you. And Zaps run deterministically, which means they're much, much cheaper. In fact, some benchmarks we show, they're about like 71% cheaper. And the reason why is because it's a way better way to ru- way more cost-effective, reliable, and efficient way to run most workflows versus having an agent have to wake up every time and reason over the instructions every single time. So y- you know, that's

  14. 14:00 – 16:03

    Model routers and multi-model ensembles

    1. WF

      one way you can sort of reduce your overall, um, token spend. The second thing you can do, and this is where it starts to get a little more technical, is y- you can use routers. Uh, and so, you know, a lot of folks will these days, um, have a frontier model route tasks to a cheaper model. So they might send it to like Haiku or, you know, an open source model or, uh, a decision model, uh, like Jev or something like that. And, you know, as a result, that can help them reduce the overall cost of the workflow on tasks where they don't actually need frontier intelligence. Of course, the trick with that is that for a lot of less technical folks, they go like, "Model, model routing?" Immediately just sort of they like nope out of that discussion. They're like, "That sounds complicated," right? And so this is why I think the best AI products have like a real wedge is to help with like this cost optimization de- decisions. Like how does the product actually just do that for you? And this is where I actually do think there's a fair amount of innovation that entrepreneurs can have because they're not incumbent to run these tasks on one model family. You don't have to just run it on the Anthropic models or just on OpenAI models. You can use a whole ensemble of models to reduce the overall cost of running the application and, um, in some cases not use models at all. Like you can write code. You can have deterministic tasks go. And so, you know, there's these real advantages that you can get when you are able to use models from a whole ecosystem of providers. And so a- again, that's another place where I think you can, you know, entrepreneurs have an opportunity to build products that are advantaged over what the labs can provide.

    2. CV

      Yeah, I, I would say probably not just entrepreneurs, but even the large enterprises. I'm, I'm curious to know-

    3. WF

      Oh, 100%. Mm-hmm

    4. CV

      ... how you're going about getting into those type of large accounts, because traditionally Zapier was meant for like the indie hacker, right? That was automating-

    5. WF

      Mm

    6. CV

      ... a bunch of things, and now you're targeting a bunch of other larger enterprises. So kind of what's the use case there for, for the folks who maybe do not have all the

  15. 16:03 – 16:40

    Getting into large enterprises, not just indie hackers

    1. CV

      decision-making power?

    2. WF

      Yeah, I mean, we spent a lot of time, uh, you know, a few years back basically working to, um, address these like gaps in our underlying key platform. Like Zapier very much was like a solo user product. It was really best in a, a small business. And so, you know, we spent a lot of time launching, you know, a lot of the sort of checkbox-y features like, hey, you've got SSO and SCIM and audit logging and all that sort of stuff. So a lot of that got addressed, but the real thing was getting really good at the thing that makes this product category scary. So for a lot of businesses or a lot of products,

  16. 16:40 – 18:54

    Turning connector governance into the moat CIOs trust

    1. WF

      you kind of have your checkbox-y features, but then there's usually one or two features that make a CIO or somebody like that go, "Ugh." You know, like if you're, if you have a meeting recorder, it's like, well, how are you handling, you know, sort of privacy? [chuckles] And like all of a sudden that's like, ah, like I get that you love this product, but like there's some, some dimension of it that like, you know, fr- freaks the CIO out. Well, for our product category, uh, that's all these connectors. You know, if you connect these, where's the data going? You know, who's, who's getting a- able to send it where? All of a sudden that makes this product category one that's like, oh, like it's cool that you can do this, but also like you might shoot yourself in the foot or like this whole business in the foot. And so we spent a lot of time building basically like governance systems on top of the connectors to make sure that, you know, if you're a CIO, if you're a leader in an organization responsible for how data moves through an organization, you now can actually have incredible visibility around who has access to what. Can you turn on read access to this system, write access to this system? The permissioning level in Zapier is even more fine-grained that you often get out of just like OAuth natively. And as a result, it makes, it takes this thing that's like a scary proposition and says, "Well, hey, if you actually standardize on Zapier with your connectors, you're gonna have way more ability to see and govern how humans access this data, how agents access this data, than you would if you tried to use any other capability out there." And so that to me is like how you flip what might- ... seem like a scary thing in a, in a product category into a strength. And so we invested a lot in making sure that that is really strong. And so now we hear from, I mean, I can think of, like, very large Fortune 500 company that went from saying, like, "Ah, we're kind of nervous about this," to now they say, "Hey, we do not let any connector or automation run on any system other than Zapier." Uh, and that even goes for, like, native integrations and stuff like that, 'cause they just trust that they have way more visibility into what's going on when they use something like, uh, when, when they use our platform versus something else. And so it, to, to me, that's, like, what it really took, was to try and understand not just the check the box stuff. We, we're great at all that stuff, too, but how do you actually address the thing that really is, like, the fear, um, that an organization has in your

  17. 18:54 – 21:10

    Why enterprises are stuck between turning AI on and shutting it down

    1. WF

      product category?

    2. CV

      I, I think that's the next big frontier in enterprise. So a lot of, a lot of players are trying to become that, call it control tower or, uh, agentic, uh, interface to connect all the different applications. 'Cause you're right. Today, a lot of these enterprises are still using AI assistants for basic use cases such as, you know, fixing typos. But the actual connection in between other tools to automate work is still something that is not happening mostly because of fear.

    3. WF

      Uh, yeah. Spot on. And, uh, you know, I, I, I do empathize with the leaders in this position 'cause you're somewhat caught between a rock and a hard place, where your options feel like you, you really feel like you have one of two options. One is to, you know, go, like, turn all this stuff on and let your team go wild building all these capabilities. And, you know, in the short term, that feels very powerful. Like, you can do all this incredible stuff, but you're opening up all these risk vectors that eventually that's gonna bite you on the butt. Or you say no to all this stuff, and you basically shut down innovation in your organization and, you know, nothing gets done. The best people who want to be the most AI forward start leaving. And so you're kind of, like, stuck between these two worlds where it's like, well, neither of these feel particularly good. And I think at, like, some of the companies struggling the most with this, it gets even harder because this stuff gets caught up in a committee. And so now they sort of sit around a room talking about it versus actually doing anything about it. And so it's really a difficult, you know, position that I think a lot of, um, you know, leaders who are sort of tasked with solving this problem find themselves in. And so I do think this is where the application layer can really solve some problems for some folks and say, like, "Hey, we've got a mechanism that allows you to put common sense, smart governance tools in place that allow you to get the best innovation that you can on, on top of it and close, like, the worst risk vectors that you might see for it." And so you kind of get this, like, third option that just is like, ah, that's kind of what folks have been looking for. And, and so that, that's where we've sort of put a lot of our effort is to sort of find that mechanism where someone can come in and say yes to a lot of people being able to build and create automations and, and, and do a lot of the things that they want their employees to be able to do but shut down the, like, most risky behaviors that tend to pop up.

  18. 21:10 – 21:34

    The discovery problem: which tools to use for a task

    1. CV

      So if we assume for a second that we're in an organization that already figured that governance out and now people are empowered to start creating some of these, um, agent workflows, a challenge that I see is discovery, right? They might not know what tools to use in order to accomplish a certain task, right? So how do we go about having AI help you figure

  19. 21:34 – 24:17

    Agents that watch you work and recommend automations

    1. CV

      this out?

    2. WF

      I think th- theirs is a really interesting area, um, to work on. I, I... One thing I've noticed for myself, a- and, you know, I think f- folks should stay, stay tuned on this dimension 'cause I think we'll launch some capabilities in around this, but one thing I've set up personally for myself is I've got agents that are watching me work now. And so it goes and looks over all the logs that I'm emitting from, you know, my agent sessions from Slack, from Gmail, from Calendar, from Docs, from code, from data, et cetera. So it's literally seeing, like, here's the tasks that you're doing every single week, and then once a week it wakes up and says, "Hey, here are some ways I could have helped you this week." And we've worked on, like, automation recommendations for years pre-AI, and honestly, like, we never really could do things that were, like, all that good. They just, the, the recommendations, they felt kind of sterile, um, and bland, and as a result I, I don't think they ever really worked. And you know, I, I remember it was like, it was always kind of like the common joke, like a new product manager or a new designer or so- new somebody would come in and be like, "Why don't we work on the recommendation system?" And, and honestly, I probably had become a little jaded about it. Like, I was just like, "I don't think it's gonna work." [laughs] I just have never seen it work. And so people just kind of, like, tech- kept taking their swing, kept taking their swing, kept taking their swing, and it just didn't work. But this agent-first version of it, like, literally the hacky version on my desktop was better than any of the, like, sophisticated versions that we had built in sort of, like, the no-code era. And so I think there's, like, a real opportunity here to actually have agents be the ones that start to make recommendations for you. Now, it has some of the, the challenges that we've been talking about. Like, you have to be comfortable with giving, like, agents ability to, like, watch all the stuff that you do. And so that's gonna, like, set off all sorts of, like, you know, flags for a, a, a company to say, "Well, how do we actually launch that capability inside of an organization safely?" And so I, like, that's where I say, like, "Hey, stay tuned." Like, I think we've got some ideas coming on this dimension but, you know, if you're, if it's just you, I would say, like, literally just go prompt your agent and say, like, "Hey, look at all the work I've done over the last week and just propose, propose automations for me. Propose things you could have taken off my plate," and see what it suggests. I think you'll find that even that alone gives you some pretty good recommendations for where to start.

    3. CV

      Honestly, that's what I've been doing for myself, right?

    4. WF

      Mm-hmm.

    5. CV

      Like, the classic AI meeting note taker and then the agent of choice scanning through my docs. But it's just for me. I can't wait to have an, you know, a more holistic view around how people are doing work, not just to spy on them, but to actually help them do work better.

    6. WF

      It's remarkable how good it works, like even just the basic version of the prompt. It's, uh, like this is the kind of stuff where you do just see the models just get better and better and better.

    7. CV

      Yeah.

    8. WF

      They're, they're incredible. [chuckles]

  20. 24:17 – 24:29

    Wade as a builder-CEO: structuring the day

    1. CV

      So I want to talk about you as a, as a CEO builder, right? That your company is really big and you seem to be still in the, in the weeds building with the team. Tell me a little bit more about how you structure

  21. 24:29 – 26:29

    No meetings before 11am, and protecting maker mode

    1. CV

      your day.

    2. WF

      I definitely have gotten, I think, better at it, or I've found my sort of like groove for how it works, in part because I just found that like as the organization scaled up, I th- I fell into kind of the common trap of just being like in meetings back to back all day, and then getting to the end of the day and thinking like, "What, what have I actually achieved?" It's like I've talked to a bunch of people, and I've sort of delegate, uh, ostensibly I've delegated out a bunch of work that is like now going to get done some way some how. But it just kind of left me feeling like not incredibly satisfied. And truth be told, like a lot of the things, we have a lot of talented people in the company, and they could've just done those without me. It's like they didn't really need me for that stuff. And so, you know, the thing I started to think about, and I, this happened after I had kids, where I was like, "Hmm." It, it was both kids and like the, you know, ChatGPT relaunch that sort of make me think like I need to like, I need to be just spending a lot more time in maker mode, because I don't think I'll be able to stay up to date on like what this, like what the models are capable of, what's the, you know, application trends, et cetera, if I'm not a, like spending time building myself. And so honestly, I just told my assistant, I said, "Hey, like I don't want meetings before 11am." Like, I'm just not gonna do that. And so that way I tend to do my best heads down time like first thing in the morning, and so I usually have a couple hours in the morning where I'm like just playing, building, thinking, mostly building, [chuckles] mostly prompting a lot these days. Uh, and that helps me get a feel for like where are these things good, where are they not so good at, et cetera. And y- you know, I, I don't know that, like it's, it's not a sort of panacea. Like this stuff is moving so stinking fast that I still constantly feel like I'm not as up to speed as I would wish I was. But it is helping. And like I can't imagine, you know, how out of touch I would be if I sort of kept operating in that sort of like back to back meeting mode all day every day. And so yeah, I think just carving out like dedicated maker session every day for a couple hours in the morning has been like the big unlock for me.

  22. 26:29 – 29:38

    Fighting the organizational urge to add meetings

    1. CV

      What other, uh, org design changes have you made in order to allow other managers to get the same type of feeling?

    2. WF

      This is very hard to do, I have found, at like a certain amount of scale. There is-- I, I don't know what it is about humans or organizations, perhaps organizations I should say. I, I don't think it's humans that do this. I think it's like just the organizational tendency, is to just add and add and add more meetings. And we're very bad at subtracting, subtracting, subtracting. And so I do feel like I kind of have to wake up every day and kind of like fight inertia and say like, "No, we don't need that recurring meeting. No, we don't need that stuff," et cetera. And so we've tried like various practices over the years where, you know, we put people on meeting quotas where you only get a certain amount of meetings a week. Uh, you know, I, you know, one-on-ones are one of those things I've certainly adjusted, where I used to do one-on-ones every single week, and now I rarely do one-on-ones with my direct reports. I think that works better for CEOs though, 'cause you're mostly working with like highly competent, highly skilled folks who don't need a lot of like week-to-week career coaching and can problem solve like in a team meeting or like just async in Slack really effectively. So I don't know if that works for every role, but certainly for my role that works really well. You know, we've definitely done kind of like some spring cleaning stuff. We never have gone as far as like Shopify. Shopify I think at the beginning of the year, they like nuke all internal re- recurring meetings, and then just say, "If it's worthwhile, like it will come back." We do like a tiny version of that. [chuckles] Uh, I think we might... Honestly, we probably should try doing like the, the bigger version of it. But I think like more than any of these like particular tactics, it really is just about continuing to, you know, walk the walk yourself and just, you know, I, like I, I just push back on meetings all the time where I'm like, "I don't think I need to be at this," or, "I don't think this needs to exist," or, "These two meetings, like it's the same group of people plus or minus one person talking about the same thing, like you should combine. Like why are you doing this?" Um, and just like constantly like re- like telling folks like, "Protect your time, protect your time, protect your time," because attention is really scarce right now. And you know, if you, if, if we find ourselves like kind of stuck in like committee logjam, like we're gonna get steamrolled by somebody who is way more like AI first, agent first, and just is like, "I'll just have the agent go build this thing, and in the time it takes you to talk about it, like I'll have the whole thing done."

    3. CV

      Right.

    4. WF

      And so you kind of just have to constantly remind people that that's like what's happening here. And I think if you do that well, you, you start to see, you know, more and more like antibodies sort of build up inside the organization a- against sort of like some of these, I guess, more traditional organizational practices, I guess, for lack of a better word.

    5. CV

      S- half smiling, half crying as I hear you give this answer because I'm there still, you know, like trying to remove myself from meetings, archive Slack channels, ask people to do the same, and they somehow keep, like meetings keep popping up, you know? And I don't know what it is about it that makes people feel like fulfilled in some way, um, that, uh, I actually think AI could be an amazing, uh, helper for this, right? The same way it can look at how people are doing work and help them do more efficient work, it could look at how people are spending time and, and give them some ideas on how to optimize.

  23. 29:38 – 30:37

    How debate meetings became a powerful prompt for agents

    1. WF

      I agree. I would also say though there is this odd new thing that's happening where because of like meeting recorders and such, and how well the models do with just like lots of messy context, meetings can also like be a really effective prompt for the model. Like you can literally point, you can get on a, like we could get on a call and debate a feature and say like, "Well, I'm worried about this, this, and this," or, "I think these capabilities will be really excellent," or, "This is the, like these architecture decisions make a ton of sense." So you literally just banter back and forth- And then take the transcript, fire off, uh, an agent and say like, "Hey, go build what we talked about." And like, I'm finding that to be like, work really, really well. So it's, it's kind of this like funky situation where there are definitely some, there's this new type of meeting that actually feels incredibly valuable, which is like, like vibrant debate over how to build something or why to build something, et cetera, that elicits this really rich context for, for an agent. So I, you know, meeting, meeting culture's hard, I guess is where I sort of net out on a lot of this stuff.

  24. 30:37 – 31:45

    Closing the loop: giving agents broad read access

    1. WF

      [laughs]

    2. CV

      Yeah. No, and I think you're touching on something interesting that involves both meetings as well as other kinds of work, which is this opportunity of closing the loop.

    3. WF

      Mm-hmm.

    4. CV

      Like making sure that you are gathering enough information and then having the agent be able to fix things without a lot of oversight, right?

    5. WF

      Yeah.

    6. CV

      So how are you thinking about that, not just for meetings, but in general for the own automation and workflows that you are creating?

    7. WF

      I mean, the, the, the sort of architecture that we're finding is that the more context you can give these agents, like the wider read access to raw data, like not just summarized data, but like give it raw access to, you know, the database, the code, the Slack, the meeting notes, et cetera, um, the messier the better. The better results you get out of this thing. Like it's able to just sort of sift all over all of that stuff and, you know, deliver just remarkably good capabilities. I think you have to be careful about how you give it write access to various tools. Um, like you do need to be thoughtful about that still, figure out where you put humans in the loop, how you have good checks on the things it can and cannot do. Uh, but by and large, like broad-based read access to a lot of tools seems to make the agent give better outputs

  25. 31:45 – 32:40

    The self-improving company, and where this all heads

    1. WF

      at the end of the day.

    2. CV

      I had this, uh, interview with the CEO of Replit, and he was talking about the self-improving company and how we're going to see more and more companies leverage this capability to-

    3. WF

      Totally

    4. CV

      ... improve much faster.

    5. WF

      I 100% agree with him. Like I think we've seen a number of folks talk about this, like Jack talked about this at Square. But it definitely feels like you're, you're gonna be able to sort of create these like feedback loops that the agents are constantly listening to. It's, you know, wh- whether they're listening to like customer complaints on social, or customer calls, or customer support tickets, or watching like, you know, performance dashboards for like how the application is doing or not. Like all of these things are generating signal around what is working and what is not. And then if you've got these like goals set up for the agent, the agents can kind of constantly be working on improving in those dimensions. So it definitely feels like this is the direction that things are, are heading.

    6. CV

      Well,

  26. 32:40 – 32:51

    Wrap-up

    1. CV

      great. It's been a pleasure to have you on the pod. Thanks so much for being here with us.

    2. WF

      Awesome. Thank you for having me, Carlos. Pleasure's all mine. [outro music]

Episode duration: 32:52

In the uListen app, the transcript lights up word by word as you listen. Get uListen

Transcript of episode 4bekHow4jlA

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.