Zapier? n8n? Meet this NEW AI Agent Platform that Just 10×-ed My Workflow

0:00You've seen those complicated N8N workflows that are all over the internet, but then you go into N8N and you find out that it's super, super difficult to create these workflows. There needs to be an easier way, and I found it. It's called string.com. It's in alpha right now. Not many people are talking about it, and it's one of the easiest ways to deploy AI agents using, yes, prompts. You can literally just prompt it and it's going to create those complicated workflows for you. I brought on the founder. This is the tutorial where he shares four workflows. The fourth one to me is the most interesting, but we go from easiest to most complex.

0:44And I can't wait to see what you end up building.

0:55this is actually an emergency pod we got todd on the pod because he just launched string.com which is kind of my dream product todd if people stick through to the end of this episode what are they going to learn with string.com they're going to learn how to build ai agents that you know make their jobs more efficient or give them superpowers, then that's the goal. Okay. Well, let's share your screen and let's get going. All right. So this is string.com, which is an AI agent that builds AI agents. And I have kind of four examples. I spent about two hours this morning going through these. So I tried to find use cases that were actually like real and would be relevant to a solo entrepreneur or a small business owner trying to like get stuff done.

1:42And we'll just kind of go increasingly in complexity if that works for you. Cool, let's do it. So the first one I had was just, you know, build an agent that monitors Hacker News for any mention of the term MCP. As you know, we've built a lot of MCP servers. When new articles are discovered, you know, notify me in Slack. And I'll just go ahead and kick this off. But my assumption is, based on doing this this morning, that it's possible there are so many articles mentioning MCP because it's like the hot topic of the day that the first version of this may be overwhelmed with volume. So it'll be interesting

2:17Tod Sacerdoti:to see how String handles it. So the first thing I'll say, don't ask me this again, but it always gives me a plan before I go ahead. This really prevents the tool from like burning your AI credits on like going in the wrong direction. In this case, it's a relatively simple prompt. And so it should have success. But we noticed previously that sometimes, you know, the AI would go build something wrong and then you'd be out of credits and be like, this thing doesn't work. So essentially, there's two tabs of string. There's the progress, which basically tells you like what the agent is doing on your behalf.

2:54You know, it's saying like, I'm going to go ahead and build this agent. I need to figure out how to get the articles from Hacker News. I need to figure out how to integrate with Slack. For those of the audience who know PipeDream, that's the company we run. We've got tens of thousands of tools available to agents. So some of these things will be pre-built and some of them will be dynamically written with code. And that's part of this increasing complexity I'm going to show. So here we are. We've got kind of a simple agent. The trigger is basically looking for articles that match the term MCP. And it's going to send the message to Slack.

3:30It's asking me to test. And this is basically, it could fail. It could say like it's somehow just not getting the articles. Let's see if it does. So what's cool about this is you've sort of abstracted away, you know, the N8N flow charts that, you know, we're so familiar with nowadays. And it's really now just prompting using plain English to actually create that. Yeah, I actually went back to a slide that I presented in January today when I was just thinking about, you know, why we went down this direction. And it was really like, if you can get out of this beautiful mind of like these really complex node charts that like everyone gets lost in and you simplify the interface to natural language, which for us is English, I think you could make the product 10 times easier to use.

4:19And similarly, if you remove the dependency on this registry of triggers and actions, which we have one of the biggest in the world, but you sort of included cogen, then I thought you could solve 10 times more use cases. And that was basically the drive to create, um, to create string. So Todd, I'm literally sweating just because this is exciting me so much. I'm literally going to have to take off my sweatshirt. That's a good sign. You know, that's a good sign. By the end of this episode, I'm, I don't know if I'll have any clothes on. strip automation is a new game. I haven't played it before, but so this is part of the magic of the product.

4:59So what you're seeing here is it's asking me to configure my Slack account. So I went ahead and said, you know, it's my pipe dream email, and then it should ask, you know, what channel I want to send it to as a next step. But again, this is kind of like built on, you know, five years of app integrations and approved clients. So I'll just say, let's do Todd test. And so now those are really the only like config components. Um, I've obviously connected my Slack account before, but if not, I would have went through like an OAuth flow. Um, but my hope would be that, you know, once this goes end to end, we can just open Slack and we'll, we'll see a message.

5:38Um, and once it tells me that we're good, I will go ahead and open it. I was successfully sent a message. All right, let's go see. All right. So this is sending, you know, per new article. So I think it was relatively intelligent. So instead of trying to send like 500 articles, it just sent me the most recent one as a test. And of course, if I deploy this agent, it will go ahead and send every new article mentioning MCP to me. Sam Altman, the co-founder of OpenAI, I just said that it is the era of the idea guy and he is not wrong. I think that right now is an incredible time to be building a startup.

6:19And if you listen to this podcast, chances are you think so too. Now, I think that you can look at trends to basically figure out what are the startup ideas you should be building. So that's exactly why I built idea browser.com. Every single day, you're going to get a free startup idea in your inbox and it's all backed by high quality data trends. How we do it, people always ask. We use AI agents to go and search what are people looking for and what are they screaming for in terms of products that you should be building and then we hand it on a silver platter for you to go check out. We do have a few paid plans that take it to the next level, give you more ideas, give you more AI agents and more almost like a chat GPT for ideas with it.

7:07But you can start for free, ideabrowser.com. And if you're listening to this, I highly recommend it. I'm going to go back. It should say, yep, yeah, the agent is ready to run and we can deploy it. So this is kind of like a basic automation, but I think what's cool about it is like I have someone on my team who actually looks for news and sends it to our product managers, product designers to just keep them up to date. Like that's a person that's actually going and doing that. This just replaced that part of that person's workflow. Yeah. And I love starting with like the most basic, which is like trigger action pair in our language.

7:46But when I built this this morning, I immediately said, can you summarize the article? Can you do sentiment analysis? Tell me if it's positive or negative. Can you write a clever comment that I could respond to the post that shows I know a lot about MCP and have a perspective on it. Right. And you can imagine this just getting more and more complex. But again, instead of that like beautiful mind of like node craziness, you know, you're just iterating through English. Absolutely. All right. All right. You want to go to another example? Let's do it. I'm not promising that I'm going to take off my t-shirt, but we will see.

8:23Okay. So I tried to find one that was a little like relevant to maybe people who are content creators. So I said, you know, build me an agent that monitors an RSS feed. In this case, I'm just going to use the Pipe Dream blog for new blog posts and then use AI to write like a viral LinkedIn post and put that content in a Google Doc just so I can go grab it and post it quickly. You probably do this like I do. I did give it a little guidance on the LinkedIn post. You know, I just said, hey, there's three viral components. Keep it short and punchy. You know, it should be one sentence paragraphs. Of course, if I was doing this, you know, for real, I would probably spend a decent amount of time on like the LinkedIn prompt just because having done this myself over a long period of time, you know, the better your prompt, like the better the output.

9:07And I would hone that over time. But again, in this example, I'm trying to now include like a few different things, like do some some AI, integrate with another app, in this case, you know, Google Docs and go from there. So I'm going to, this is just one of the little nuances, but I've turned on auto testing. So as it builds, it'll just automatically test. So I don't have to constantly click test, test, test. But if you were doing an agent that might like delete rows from a Google doc, you know, you kind of might want to have human control over that. Okay. And what's, by the way, we never talked about pricing.

9:45Like how does this work? Yeah. So we're really following the model of the Vibe coding platforms, which is you basically have a subscription to the product and it comes with a pool of AI tokens. I believe the token count is 20 million on our basic plan. You can see I have 58 million tokens in my account right now. So we can monitor through this process how many I use doing all these use cases. but of course like our goal is to you know reduce the cost of token creation and token usage and

10:17Tod Sacerdoti:just you know as these models get better reduce those costs and i see a button for open and pipe dream what does that mean yeah so um it's funny i was uh i was in a meeting kind of demoing a very early version of this to guillermo from versell and he said you know my sort of view these vibe coding platforms is their best when they're built on top of like dev tools that have been built over like four or five years. And the reason is because you can have this escape hatch. So we have this button where you can like open this into our old product and essentially go back into that WYSIWYG editor and, you know, make changes and do the drag and drop and all that experience.

10:59I actually slacked my team when I was doing this this morning. And I said, I only had to do that once this morning. So to me, it's like, it's a little bit of an anti-pattern, right like we want users to never have to open um in pipedream and and eventually that button will go away um and i think that will be you know when we're no longer an alpha product but we're you know a successful product that button will be gone it's interesting by the way i don't know if you caught that while we were talking but um it looked like oh it's happening again looks like there's a there's a some sort of error in the trigger um i'm not 100 sure what is happening but it's obviously not receiving a trigger event.

11:41It should recover. Oh, here it is. It's explaining. So it says it's, it's not receiving data for the 25th event. So it's basically like looking for blog posts that happened. I guess that was three days ago. Clearly we need to produce more blog content as a company because there was nothing recent enough. So I assume what it's going to say is I changed the date. Okay. I've set the published after date to May 1st. So now it's like, it's going further back to looking for like the most recent blog post, which I think is actually pretty intelligent, right? It started with a narrow date range. Then it probably expanded from one day to three days.

12:16And then it's like, all right, there's nothing there. I'm going to go all the way back to like two months ago and grab your most recent blog post. It'd be interesting to see the most recent blog post was June 11th. So it needed to go back, you know, a little more than two weeks to find that. So I, again, I consider when, whenever you're vibe coding and something fails, but recovers, that's more what we're focused on today really than anything else is like trying to figure out how to get people to value. And does it fail often? And like, how should people think about dealing with the failures? Yeah.

12:50I mean, I think it really depends on the complexity of the use case, but my sort of general belief is if you're just testing these tools, you should probably get, you know, 50 to 75%. If it's something like a prompt on the homepage to be higher than that. But I mean, if you're just doing like simple things, want to see if I can get to value, like I would think over 50%. If you're doing things complicated, whether this is like, you know, replet, bold, lovable, or, you know, a product like String, you know, I think you're in that 25 to 50 % success rate today. And that's why these products are, you know, in our product leases in alpha, you know, that that's the percentage we're monitoring internally is like, how do we get the majority of people to value?

13:33So it says it, you know, created a, used a open AI in this step to create a LinkedIn post. One thing I'll mention is we have this kind of like what we call batteries included for AI. So we used to include all these AI tools, but you constantly had to like input your API keys or figure out, should I use this image creation thing or that image creation thing? So now when anything touches AI, we default to the tools we think are best for your use case. And we just include our own API keys. And then we just bill you credits when you run these agents in the future. It's another one of those examples of, can we get more people to value?

14:12So let's see if we created a Google Doc. Okay. Let's see. Make sure this is a, let's see. How Alter builds integrations faster without hiring a single back. I mean, again, I'm not going to go through all the detail on this, but like, looks like a, you know, kind of like a legit LinkedIn post to me probably would need to edit it. But, um, a summary of this, this blog article. Yeah. It's insane. Yeah. So, and if, and if you wanted to, and when you say edit, like you mean a human being edited, editing it? Well, I'm just saying when I do the final link, I don't know what you do on LinkedIn, but like when I do the final LinkedIn post, I kind of take what the AI has given me and then I, you know, I polish it.

14:56That's like the last five, 10 minutes. I want it to be my voice. And of course I'm trying to train it to get better and better. But I always think that last polish is like, is the best human use of time. So what you could, could have done here, for example, is add another step that pings a Slack channel that says the Google doc has been created and is ready for review. Right? Yep. Absolutely. Yeah. And if you wanted to do that, you would just type a message? How would you do that? Yeah. I mean, it's almost like an employee, like your AI. Everyone's hiring these AI automation engineers. It's like the new hot job rec.

15:33We thought about calling this product just like your superhuman AI automation engineer, because it's working for you at all times and available for work. I'm just not sure that that's a widely understood job yet, but maybe it will be.

15:48Tod Sacerdoti:actually. So what's it saying here? So it says the user wants to add a step to process when the process is completed. So it needs to configure a Slack notification step. So here it's just saying send Slack notification. And then I assume it's got to write that. And then of course it's got same exact flow we just went through before. It's got to figure out where you want that to go. We've tried things where we just like automatically select Slack channels. Like, so we have to be smart about it. Like there's certain apps where you can, you know, make an inference and figure things out, but it's like, you probably don't want this, you know, little test that you're working on to exactly like click the general channel.

16:26Like you probably don't want to like blast your company, you know, with all your little tests. Um, so anyways, that's, that's kind of what it's doing. It's just figuring out like where to send it. Um, and you know, I assume, uh, we'll see what it says. Let's see. Uh, it's running the test.

16:45Tod Sacerdoti:All right. Successfully test. See here. I don't see. Oh yeah, it is. Yeah. New LinkedIn post generated. Nice. It thinks it's, it's high quality enough to get a rocket ship emoji. Um, it didn't link to the document, which of course I could ask it to do. Um, but you know, essentially here it goes, it gives you the document ID, but didn't give you the URL. So, and that's a case where I think it was kind of a dumb decision. could have been better. We'll tell string to do better next time. Love it. You want to move on to another example? Let's do it. Okay. So I'm going to give an example. So I was really inspired by your boring marketer team that had like a little chat about string in on Twitter, which I saw, which was super cool.

17:28I liked it because it was a real world use case. They basically said, you know, give me a daily summary of my Google analytics. So I'm going to fire this up and then I'll talk through it. And so I would say this is an example where based on my test this morning, so I might need to jump to one that works. I'm thinking like 50 % success rate. And so already I got some ideas on how to fix this. But about three weeks ago, we sent an email. It was like to a million people and it included a URL to a Vercel app. And we basically built a demo of like, it was chat. We ended up putting on chat.pipetrain.com.

18:05But at the time it was a, a, just a Vercel URL and somehow Vercel URLs like get flagged as like really bad in Gmail and these other email services. So like for the last three weeks, I've been logging in every day and like into Google postmaster tools and trying to figure out like, are we out of the penalty box? Are we out of the penalty box? Are we out of the penalty box? So, um, I basically just said every day summarize, I should have said the last seven days of Google postmaster statistics and send them Slack so that I can break it out by day and include a summary of the week so I can figure out, you know, basically are we essentially out of the penalty box or not?

18:42It turned out about, you know, a little over maybe four or five days ago, we kind of like got out of the penalty box. But the reason why this is particularly interesting is what you can see on the screen right now is, you know, we've integrated Google's APIs, but we have nothing in the registry related to, like fetch your Google Postmaster tool. So this is all like dynamic code. And this is the second half of what I talked about, which is one, communicate with natural language instead of like a WYSIWYG editor and two, you know, get out of the registry and write all these things dynamically. And so basically string is writing custom code to hit the API to figure out, you know, stats, and then we'll ultimately format that message for Slack and then send that message to Maybe we've overused Slack in this demo, but we do live in Slack.

19:33And when I did this this morning, you know, one time it sort of veered off and got a little bit lost and it was really struggling to get the statistics from Postmaster Tools. And then the second time I did it, it had one error, it fixed it and it worked. So it'd be interesting to see, you know, how successful the test is. I mean, if this is able to work, like that's a huge, that's a huge win, right? Like that's saving you a ton of time. Yeah, I think for most, what I would say is like most people just wouldn't do it. Yeah. You know, like when I've talked to the CEOs of other no-code platforms, they always say the number one problem they have is their product is too complicated.

20:17And these are the people who run the least complicated no-code tools in the world. So I think that just the fact of the matter is if you have to go through some multi-step config that requires you to understand an API, you've lost at least half and probably more like 80 % of the potential audience for these types of products. So you already see, by the way, this is already, I already know it's gonna air because it says it's received data for zero days of the past seven. So I'm interested to see if it recovers for that. Yeah, I think like as you're going through a lot of these examples, I'm just thinking about my own business.

20:53Like what are automations I can do help save time and also just like make more money. Yeah. And I, the way, the way we like view this, we have like this worldview of these automation, you know, agents. The first is tasks. Like if you just want to send an email, send a message to Slack, add a row to a Google sheet, I think ultimately that's going to be built into the AI chat apps. You know, I think you're going to go to Claude and you're just going to be like, you know, like send a message in Slack and it's just going to do it. So I think that like, you don't need a standalone tool for that. The second category is automations, which I think we're starting in automations in this conversation, we'll kind of move into true agents, but automations are the things we all know.

21:37They're like Zaps or PipeTream workflows or, you know, whatever the tool of choice is for a company. And I think those can absolutely move to a text-based interface. I think we've sort of proven that. And then the third category is what we internally call like real AI agents. These are things that have like autonomy and access to tools and sort of don't need this step-by-step guidance as much as there might be a step that says like, you know, go do these things. And it just figures out, is there steps? Is it one step? Is it one app, three apps? And we're spending most of our time, you know, on the product side, either trying to get people to value just like the chat interface or building the foundation for these more autonomous capabilities.

22:23I mean, that's the dream. The autonomous agents is the dream, that you set a goal for one of these invisible employees and they have some constraints, but they're basically working at the problem and they're trying to get closer to that goal and they're working for you autonomously. And we really try to not overpromise those things. I feel like half these tools are like, it's solved. And I'm like, it's not solved. It's like there's a lot of momentum and green shoots and potential. But for the most part, and we have another half of our business, by the way, where we sell infrastructure to other AI agent companies.

22:59And we see there's a lot of promise, but it's early. It feels like the first out of the first inning, not even the first thing. So anyways, this one just said summarize Postmaster stats on the Slack. it actually had one test failure and then it actually got to success. And I said, Hey, the key metrics are blank. And then it went ahead and fixed it. And then, you know, it's sort of explained to me like the visual representation, you know, basically said, you know, here's your results for the last week. This was eight 30 this morning. All the metrics are good, except for this one, which I know, which was the one where we got hammered on our reputation, did a little bit of analysis and then said, you know, focus on improving sender reputation.

23:45It was kind of like a, you know, no shit. Um, I got to work on that. So, uh, but that was the, uh, that was the example that worked this morning. You know, this gets me thinking that like productizing agents and automations, and then kind of wrapping, wrapping that up and selling that to other companies is a great business to start. You know, I'm talking about like a cash flowing business. Like if you're, you know, I'm sure someone would pay$29 a month for that service set up. So have you seen people basically start using string to create some workflows and agents and then selling that to, you know, other companies?

24:24Yeah. I mean, we've, we've talked a lot about, I think it was a little like, at least on our platforms a little early, but I think if you look at what's happened with N8N and all these, you know, they call it the, like the automation bros on like Twitter and YouTube, you have this demographic of people that, you know, maybe we're in, you know, drop shipping and then we're in crypto and then are like, they're in the business of making money on the internet. And they've sort of discovered this like AI agent category as a potential market to like, you know, build products and sell them. At least from my conversations with customers, I think most of those is more about today, at least generating views of their content and actually less about selling like quote unquote the templates.

25:05Um, but I think that there's a huge opportunity once these templates like are actually easy enough to implement for the end user that like, I think there can be really big businesses built in these. It reminds me a little bit of like the kind of like the website template world where, um, you know, the challenge is always, can you actually get to value if you purchase it? Right. And so these things need to be easy enough to implement for the end user. And then I think absolutely you could, um, you could build a business on that. Before we get to the, to the next one, do you have any, you have one or two ideas for people if they were trying to build a big business on top of string, what, what they, what could they build?