Overview
This episode centers on how AI is changing both software buying and day-to-day work inside companies. Wade Foster, founder of Zapier, explains two major shifts: first, that businesses may increasingly need to “market to agents” rather than just humans, and second, that while AI makes it easier to build internal tools, SaaS is far from dead because polished software still beats most homemade alternatives.
The conversation also explores how Foster personally uses AI models and custom “skills” as a leadership copilot, showing how executives can create practical workflows for decision-making, writing, and problem-solving.
Key Takeaways
One of the most novel ideas in the discussion is “agent marketing.” Foster argues that AI agents are beginning to choose products on behalf of users, which changes the rules of marketing. Instead of persuasive design and emotional messaging aimed at people, companies may need clear, structured, machine-readable content that helps agents understand exactly what the product does and when to recommend it.
A counterintuitive point is that agents often prefer stripped-down experiences: fast-loading pages, plain text, clean documentation, and highly descriptive language. What works on a polished landing page for humans may not work nearly as well for an AI system deciding which vendor to use. Right now, this field is still immature, and even the most advanced practitioners are learning through heavy experimentation rather than fixed best practices.
On the SaaS question, Foster makes the case that AI-assisted internal development does not automatically replace established software products. His CTO was able to quickly build a meeting-recording tool as a proof of concept, but the company has no intention of canceling commercial subscriptions. The reason is not technical feasibility but economics: software spend is relatively small compared with headcount, and maintaining internal tools diverts engineering effort away from core business priorities.
Another expert insight is Foster’s use of multiple AI models as a “war council.” Rather than relying on one assistant, he switches between models to get different perspectives, especially when solving hard problems. He finds Claude more conversational in many cases, while OpenAI’s Codex can be better at debugging and critiquing stuck work. This suggests that effective AI use is less about picking one perfect model and more about orchestrating several.
Practical Steps
If you sell software, start testing for agent discoverability:
- Create a lightweight, text-first version of key product pages.
- Make documentation explicit, well-structured, and easy to parse.
- Track whether models like ChatGPT or Claude recommend your product for relevant queries.
- Run repeated prompts and compare how often your tool appears versus competitors.
If you lead a team, use AI as a structured copilot rather than an ad hoc chatbot:
- Build a central folder or knowledge base with strategy docs, meeting notes, decisions, and recurring workflows.
- Ask the AI to organize that system for you instead of manually designing it.
- Create reusable prompts or “skills” for recurring decisions, such as hiring, prioritization, or communication.
For technical work, use model comparison intentionally:
- When one model gets stuck, ask another to critique its work and fix the mistakes.
- Give the system direct access to relevant tools, such as email, calendar, or Slack, through connectors like MCP.
- Be explicit when needed: tell the model which tool to use and what action to take.
Notable Quotes
“You're no longer advertising to the human. You are trying to get the agent to say, pick me, pick me, pick me.” — Wade Foster
“For load-bearing infrastructure inside of a company, there's not a lot that we're building and replacing.” — Wade Foster
“I want to build out a system that makes me amazing as a CEO.” — Wade Foster
Full Transcript
Are you marketing to agents now? The agent is actually choosing what products to buy on behalf of a human. And so you're no longer advertising to the human. You are trying to get the agent to say, pick me, pick me, pick me. Is SaaS dead? Our CTO vibe coded like a meeting recording software. It's a cool proof of concept, but we're not canceling our subscriptions anytime soon to our meeting recording software. You have a Claude skill that helps you lead. He calls it his war council. What this skill is, is if you have a particular problem or a decision you're making, you can just say, hey, I want to invoke the war council to like advise me on this. Do a screen share. Let's see how you work. Sure. Okay, here we go. Wade Foster is the founder of Zapier, the AI automation company. You told me about someone who you hired recently who impressed you this agent marketing guy. What's agent marketing? Well, so I think this is a new thing that folks on the cutting edge are paying attention to, which is the agent is actually choosing what products to buy on behalf of a human. And so you're no longer advertising to the human. You are trying to get the agent to say, pick me, pick me, pick me. And that is often a very different skill set than marketing to a human. You know, with human, there's all sorts of like advertising history around like consumer psychology and like how do you position these things? And the way an agent makes a decision is somewhat different than how a human makes this decision. And so, you know, this person is building all sorts of tools to like track, like, you know, when does ChatGPT recommend this tool or not? When does like, you know, Claude recommend this or that? And why do they recommend it? And then trying to figure out how do you like, you know, structure your content, structure your APIs, structure your CLI tools such that, you know, more often than not, it's going to say, oh, when I have this problem, I want to go use that vendor versus that other vendor. You know what? I had that experience. I used Cloud Code and I said, okay, now how do I publish it? It said, I can just publish it to Vercel. I go, I never heard of Vercel before, but okay, go ahead if that's what you can do. And it did it. And then a few days later, I was at a Cloudflare event and they talked about how they can do the same thing, but better. And I said, wait, I would have preferred all those features, but I had no idea you even existed. So what are you finding that is effective for marketing to agents? Well, so agents care a ton about like clean documentation, like fast web pages. So, you know, it's like the thing I'm seeing a lot of folks do is serve up a separate version of their webpage that is meant just for the agent. And so that content is like plain text. It's served up like really fast. It's like written, like very, it's like very descriptive, almost like mechanical, because like, that's what the agent prefers. It's like, I want to have just like, give me just the facts in like an easy to consume way. Very clear, very direct. And as a result, like it, you know, it, it's able to like understand what you're offering much better. Whereas like, you know, a human cares a ton about the design, like the visual design and things like that. Agent, it's not so much, right? So it's a different, you just kind of have to like figure out like, what is the thing that it's actually optimizing? And I'll be honest, like a lot of this is, you know, very much like a more art than science right now. Yes, people are building evals and, you know, things like that to kind of track it, but yeah, when you ask like people who are really good at this, their answers are like very hand-wavy still. So what did he say or do that made you think, okay, he's got this? Well, I think just the level of experimentation is just much, much more. That he's experimenting so much. Exactly. But how do you even figure it out, Wade? You do a test. It takes a while for the agent to even know that you existed, let alone to know that you've made the change. What do you do to test and know quickly? I mean, they're just like running a bunch of queries. They're having their friends run a bunch of queries. They're trying to pay attention to these things. There's like automated techniques that they're using. Yeah, so there's, there's, you know, it's, it's like a lot of trial and error. It's a lot of trial and then going in and doing a query and seeing what the response is. And then there are tools that you use to figure that out. We did a, I did an interview, frankly, with people on your team who talked to me about some of the software that you use. We'll have it up on the channel for people to see that. Okay. So it's trial error, keep measuring and seeing the results and start to get some clear understanding of how they think. And one of the things that you said is clear, a lot of documentation, clear writing, serve up even a second page just for the agents. What else? What am I missing? You know, those are, those are kind of the, the, the big things that I've heard of right now. Wade, everyone is talking about whether companies are going to be buying SaaS or are they going to be building everything internally? You told me about someone on your team who built something really cool internally that you could have bought. What is it and who is it? Well, our CTO like vibe coded like a meeting recording like software, you know, sort of like a, like a Fathom or a Granola or, you know, a Grain or something like this. And, you know, it's, it's a cool proof of concept, but we're not canceling our subscriptions anytime soon to our meeting recording software. Why not? Why not? You can build it. You can save yourself some money. You can't, you can't like the maintenance on this stuff, like the tokens you burn setting it up. Like, you know, at the end of the day, like these products are not that expensive and the teams behind them are just putting so much effort in polishing them, making them better, advancing their capabilities. And those are all things that we would have to staff engineers to keep working on them. And yes, it's gotten so much cheaper to build things, but I think we would rather deploy that engineering power to making our stuff better versus, you know, improving our internal tools. You know what? I have an example of that. I built myself the ideal calorie counting app and wait, it is so perfect. It's like right on my watch, which is the app that I use all the time. And still there's always like an edge case random thing that comes up that I have to go and handle. It makes me realize, oh, of course that company already thought of that. That's what you're talking about. You don't want to divert your people. Yeah, exactly. It's like for, you know, for, for, I think vibe coding is fantastic for building tools that you can't buy or these like, like smaller utilities where it's like, Hey, this is core to sort of like a one-off piece of software that I need to use to solve a particular job. Fantastic. We're seeing so much more of that stuff happen, but for like load bearing infrastructure inside of a company, There's not a lot that we're building and replacing. Like there's a lot of really good software out there that we're very happy to continue being customers of. And overall software is what would you say? What percentage of the business spend? It's minimal. You know, it's, it's like single digit percentage. It's so small compared to like headcount. Essentially really, that is what it comes down to for enterprise. It's one thing for an individual like me to say, I don't want to spend money on a calorie counting app. That's going to charge me 20 bucks a month. It's another thing for a company to say, it's, it's just a small price and a big distraction. Yeah. It's a drop in the bucket of, and another thing to maintain support, pay attention to, et cetera. Let another company who cares about that deeply be awesome at that thing. You know, let me be honest with you, Wade. I'm kind of shot out of a cannon right now. Like this energy is a lot of high energy and I'm trying to restrain it, but I have to be honest with you, even when I'm off mic, I'm talking like this all the time. Hey, Olivia, check out what I could just do right now. This is amazing. Look, my cloud bot just did this and that. And she's like, I get it. Tone it down a little bit, but I can't, I'm so freaking excited about it that sometimes, and honestly, it clouds my judgment. Before we got started. First of all, do you feel any of this or is it just me? Well, I'll tell you what I built. So a guy who runs all our social stuff built a skill that's going viral inside the company right now. It's, it's something you've been able to do for a while. So it's not actually new, but it's the, the, the way in which he deployed it is somewhat novel and he calls it his war council. And so what this skill is, is, um, it's spin, if you have a particular problem or a decision you're making, so say like you're trying to, uh, decide on hiring a new person or say I like using Cursor because you can switch between a bunch of different models. So you can see here, I have like Clopus, Claude 4.6, but I can switch into, you know, 5.3 Codex, Gemini 3, et cetera, like very fast, which I find to be pretty useful. And, you know, you just have this like chat tool here. But before you do, what are you finding that's more useful in one versus the other? So I think the, I find switching is really useful when you're working on a problem and you want to get, again, different perspectives. So like, I'll have it critique its work. One thing I've found is that like Opus is like more enjoyable to just like talk to and go back to it. Like sort of, it's an easier conversationalist, but like when I'm having it work on hard problems, it sometimes just like bugs out and doesn't do as good of a job. And so oftentimes for those, I'll switch over to Codex 5.3, which is the OpenAI model. And that, again, it's not as good conversationally, but it is just like much, much better. And so like I was dealing with this like annoying bug with the, I had to build this like old MCP server for a tool that I was using. And Claude was just like, it just, it just like kept, I just kept going and going and just couldn't seem to get it right. And so I toggled over to GPT-5.3 Codex and I said, hey, critique the prior agent's entire work and tell me where it's getting stuff wrong and then go fix it. And it was like, it's making this dumb mistake. I've got you. Oh, that is such a great idea. You know what? Even when I, with my little programs, there are times when it's just stuck and it can't solve it. To be able to go say, hey, go critique, figure out what it's missing. Okay, brilliant. Okay, so now I see your setup here. I see the beginning of it. What else? How else do you get context into it? Well, so, you know, basically you point Cursor at a file system, which, you know, the file system is just like files on your machine. And, you know, I usually keep the file system closed, but, you know, you can see here, all the like contexts that I've got into this thing. And so, you know, I, I have all my like daily. I have it create a daily brief and a daily review for me. So you can see a whole bunch of stuff there. You know, I've got a bunch of strategy stuff in there. You know, I've got, oh gosh, this is, I can't show you this. Oh, because it has all your meeting notes. Well, it's just like some of the file names are just like, like a giveaway on certain stuff as well, too. But you can see I've got like strategy, decisions, leadership, people, meetings, initiatives. Okay. All this stuff is probably fine to show. It doesn't look at all of those, does it? To consider any one question. It just has the ability to, and you evoke it with an at symbol. Well, so here's the thing. I have, it does, all this stuff is in the context because it's pointed at my co-pilot. So it can go like crawl the system and find it on its own. But I can also say, you know, uh, hey, like, um, uh, you know, let's see what's, what's an interesting example. I'm trying to decide whether to continue this interview with Andrew. Yeah, I, I want to figure out if the, uh, I should keep working with Andrew or not. Uh, remember, so like, I could just ask it like this and it will, you know, sift through all of the context that I have shared. But I can also say, remember to keep in mind, uh, the, you know, company strategy, right? Something like that. So if I really want it to like, Hey, like really index on this, I can do something like that. Okay. And do you usually do multiple skills at once? Like, would you, or sorry, would you say company strategy and use this skill? Or are you just picking one type? So the skills you can, skills usually just have like keywords that you can invoke them on. So I can say, you know, uh, you know, should I work with Andrew or not? Uh, whoops. Uh, it probably doesn't know. We'll see. It's just thinking. We'll pick the wrong Andrew. I know there are multiple Andrews. So should I work with Andrew or not? Uh, and then please invoke the war council. So like that, I didn't have to do anything fancy. It just knows that I have a skill in here somewhere that is the war council. And I actually don't know where it is in this system, but it's in here somewhere. And, you know, the, the reason, um, you know, like, there's sort of like, like the, the fancy thing about like that I see a lot of people doing is, you know, a lot of people will get caught up on like the file system. They're like, how am I supposed to organize all that stuff? What is that all over there? It's like very confusing. Honestly, it doesn't matter. Just like ask the agent to like build all that stuff for you. And so, you know, if I'm building out a skill, I'll just say like, Hey, make me a skill for this and put it in the skills, like put it where you think it goes. And so it just organized. It keeps everything organized for me. And when I, when I built out this whole structure, the way I built it out was I said, Hey, I want to build out a system that makes me amazing as a CEO. Like help me build out, like create you to be an amazing co-pilot, amazing, like chief of staff for me as a, as a, uh, assistant and build out a whole set of like markdown files and, and files and systems and all that stuff that like helps you help me. And you should get better as we work together more. And that's all I told it. And so it organized that entire file system for me. I don't really look at it all that often. It just, it's just there. You know, my one frustration with all this is that it's not in the cloud. And I know that that's a benefit too, but sometimes I'm on a different device or I want to be in, I want to be on my phone. Totally. Okay. What, one other thing that I'm wondering about with your setup is do you use MCP like Zapier's MCP to be able to, for example, draft an email? So once you come to a conclusion, you will. How do you invoke that? How do you say, use the Zapier MCP and draft an email for me? So you, um, you just can turn on like a bunch of, um, uh, you can create your Zapier MCP server and then you add tools to it. So you could add like Slack, Gmail, calendar, you name it. And then you don't usually, you don't have to like, again, you don't have to do any fancy stuff. You can just say draft me an email, please add, add this to my calendar event. Now, if you do say use Zapier MCP to draft a email inside of Gmail for you, you will find that it like definitely gets it right. Sometimes if you don't like, if you're not that like verbose with it, it will forget that it has access to use those tools and you'll have to remind it and say like, Hey, remember you have access to these tools. Where do you give it the tools? Uh, where? Yeah, where? Um, so you can go to MCP.zapier.com and just start turning on tools. And then how do you tell, uh, Opus 4.6 in this case or, or, uh, Codex, how do you tell them that you've got, that they have access to it? So they will know because you authenticate, uh, your clients with Zapier MCP. So on Zapier's site, you press the button, you connect it into Claude. It now has it forever and ever. Exactly. Just like if you were connecting like your, you know, authenticating your Gmail account to a tool or your Twitter account to a tool or whatever, it's like big window pops up and says, I grant access. And then boop, it sort of can go work with it. Honestly, I found it's so much easier. We can stop sharing soon. I found it so much easier to connect it to Claude than I did to connect it to, uh, to ChatGPT. Like the other day I just said to Claude, do you have access to this? Can you do it? It goes, yeah, you gave it to me. And I give it to you, gave it via Zapier MCP ChatGPT is kind of frustrating that way. All right, you've got. Yeah, they're all kind of at different like, uh, parts in their journey on how good they are at working with connectors right now. You told me a few months ago, I'm not a developer, but I'm getting really into Claude code. You've gotten into it. Now you've got a team of how many people working at Zapier? There's 800 people at Zapier. How do you get them to feel this level of it's not me, but I'm gonna try it. And ideally the level that where you are, where you want to go say to Chelsea, you got to find out. You