I Built AI Agents to Replace Our Marketing Agency (7 Meetings Booked)

At some point, every marketer pays an agency to do something they know they could do better themselves. I hired one for outreach. They sent 100 emails, got three replies, and two of those were people asking us to stop emailing them. The problem wasn't just the results. It was the weeks of waiting, the wrong people getting targeted, emails that clearly weren't written by anyone who understood what we actually do. But AI has changed what's possible here. It doesn't get tired, it doesn't cut corners, and unlike humans, I can speak freely without hurting its feelings. So I'm building a fully autonomous agent pipeline that prospects, vets, and writes personalized emails to complete strangers convincing them to tell me their greatest [music] marketing secrets. And if this fails, then I wasted my time and we're back to paying [music] thousands for outreach that barely works. But before I could build anything, I needed to figure out what I was actually going [music] to build it in. I narrowed it down to three options: N8N, Claude, or Open Claude. N8N I ruled out pretty fast. So it kind of just feels like Zapier, like you're just dragging and dropping steps into a visual workflow. Like it's cool, [music] but it doesn't really feel like I'm doing anything AI native. That left me with Claude and Open Claude. They're both really good options, but I like how with Open Claude I'm not tied down [music] to one model. If I want to generate images with these emails, then I can easily plug [music] in Nano Banana. And so it's just a lot more flexible. But what sealed the deal for me was something pretty simple. Open Claude can run through Slack. Okay, so I'm on Slack like all day and there's just something about working [music] inside the tool that you're already doing work in that makes it feel so much more real and productive. [music] If I do it in Claude, I'd have something like this super complex agentic workflow, and then here, what do I have? I have a recipe and then like symptoms of stomach pains [music] or something below that, it just feels kind of weird. So, with Open Claw installed, locked [music] and loaded, it was time to actually build the thing. And it all starts by giving a system taste. The thing with AI content is that it all kind of just sounds the same. And if I just let it loose on outreach, it's going to feel like spam for the people who are getting these emails. So, the voice, the taste, [music] the judgment, that has to come from me. The point of creating this, it's not just to have AI do the actual work, it's [music] for AI to do it the way that I would do it without me doing it. Do it! Okay, so here's the plan. I need a project manager first. I'm going to call it the campaign coordinator. Then we've got three specialist agents underneath, a prospector, a vetter, and a copywriter. The ones actually doing the grunt work. And to make sure the project goes how I'd actually do it, the coordinator's [music] going to interview me. Like, to the point where it's asking me an annoying amount of questions until we're [music] completely aligned. And if it works the way that I'm envisioning, it's basically foolproof. To create this interview, I needed to make a markdown file for the campaign coordinator, which is basically text [music] instructions that tells the agent how to do its job. I used Claude to generate this file. [music] Then I repeated the process for the prospect, the vetter, and the copywriter. And when I went to review the files, something was clearly off. What the heck? After tons of review, I don't know what that means. multiple fixes, Give me a bullet point list of all the fixes [music] required. There's too many. And another solid hour of work, corrupted file. Regenerate it. We got it to a good place. Okay, so now it's generating the instructions for my main agent Abot. Now we get to actually test this out. I'm pumped. It was time to find out if this thing actually works. I opened Slack and prompted my Slack bot to kick off the campaign. Come on, do it. Do it. Start asking [music] me questions from the ICP interview. Success. This is sick. Okay, I hope this works now. This is the easy part. So, I went off to answer the questions from the ICP interview starting with why I'm even doing outreach. We're launching a new video series called Traffic Hunters where I sit down with founders and marketers who've gotten exceptional results using digital [music] marketing and I asked them to share their juiciest secrets. For this outreach campaign to be successful, I need to book at least five meetings. Now, all I had to do was tell OpenClaw the ideal guest profile. I think a good fit are websites that are ideally not super authoritative and that they have significant search [music] traffic. Still going through the questionnaire. I don't know how long this is going to take. If [music] it works, it's all worth it. A few more questions on vetting criteria and personalized templates and bing bang boom, we were done. All right, so here we go. Perfect. Let me draft the template based on what you've described. That's decent email. Done. That's a wrap. With the campaign brief completely filled out, it was time for the moment of truth. Is the campaign coordinator going to spawn the prospector? All right, come on. Come on. Please work. Please. Man, this would be so dope if it actually works. Oh, and I forgot to mention I built a dashboard so I could track everything that was being done. Nothing yet. Oh, jeez, it's working. Okay, so let me walk you through everything that is happening right now. We have the prospector that's going in and looking for for contacts using the Intros MCP server. The campaign coordinator created this outreach campaign here basically saying that we're looking for 200 prospects. We have the vetter, the copywriter, the drafter, [music] and as they go, they're supposed to check the different items that happen. And then we have [music] the actual outreach tool that was created. So, in the settings, they had contact title priorities, [music] and it's going to look for people with those titles basically as a first contact point. And then we have our template, which was generated and approved in Slack, and now we're at the point of adding prospects. I click back to check on the campaign, and nothing could have prepared me for what I was about to see. Oh, jeez. It's adding prospects. Oh my gosh. Vacuum Wars, man, that sounds like a cool niche. Holy crap, they've had some serious traffic growth. Oh, and then it's going down. We're almost at the end here of the prospector. Let's see if it actually marks [music] the task as done. Yes, it's done. Sick. Found 200 prospects. [music] Let's see if it sent the message in Slack. Nice, prospecting's done. I found 165 solid prospects across food, travel, and fitness, and sports niches, all enriched with Ahrefs metrics. 124 have emails, 89 have named contacts ready for freaking vetting. Yes. Then the campaign coordinator spawned the vetter, and that's when things started to unravel. The vetter's done apparently, and there's no way this is right. The vetter had disqualified almost every single prospect, including sites with clear consistent traffic growth. This is growing traffic. How is that declining? So, we need to find out what happened here and fix the vetter in the way that it's looking at traffic. After some debugging, we found the problem. The traffic slope field in the enrichment data is calculating incorrectly. It's showing negative values for sites [music] that actually have massive growth. Looking at your screenshot example and the raw data, prospects with clear 5-year growth are showing negative slopes and are getting disqualified. Okay, so it's a it's a bug in the software. Bug in the tool. Fix the core issue now. More hours of debugging and finally I think things are better now. So basically it said this is a permanent fix. There's 102 qualified now and 63 disqualified, zero flagged. Sweet. I think it's doing it properly. Moment of truth. All right, [music] let's refresh. Oh my gosh, I think it's actually working. Yeah, that's a growing traffic trend. The tool got it right. All right, so I'm going to go and I'm going to run the copywriter one last time which is kind of our final test. We were almost done. Oh my gosh. Sweet. Wait, why is it doing that though? NO! It's not actually merging the freaking things there. Oh my goodness. That's a a big problem and this is why we don't automatically send emails for that very reason. Yeah, it screwed up royally on my drafts here. All right, so a sub agent's gone and working on all the fixes. So now I'm just debugging. It's waiting game now. All right, let's watch it do its thing. There's six tasks. I want to drain six pots before those six tasks are done. But before we find out who wins, let me tell you about the irony of this whole project. Well, okay, we're one for one. I spent days building a custom agent pipeline to do prospecting, vetting, and outreach. And the entire [music] time Ahrefs has been cooking up Agent A which can do all of this natively. [music] Oh, not a chance AI. That was perfect. It's an AI marketer with unrestricted [music] access to every Ahrefs metric. Prospecting, done. Vetting traffic trends, done. Connecting to my custom outreach tool, yeah, it can do that too. Done, game. Oh, no it's not. Shoot. Shoot, it's at five tasks. [music] So, if you want the version of this that doesn't require six hours of fixing slope calculations, link's in the description. All right, final putt. Six fixes. Let's freaking go. I got it. AI can fix six issues with my outreach system almost in the same time as I can drain six putts from like eight feet away. That's wild. With everything finally working, there was only one thing left to do, [music] run the whole pipeline autonomously. But, I had an early morning flight to catch. So, I packed things up and the next morning I would put everything I worked towards to the test. [music] Okay, so I got to catch a flight in like 40 minutes and I'm about to try to run this in full autonomous mode. I don't know if it's going to work or not. Let's flipping go. All right. Autonomous mode activated. Prospector spawned. Okay, we're starting to add people. Done, done. Oh, no way. Prospects started to roll in. Time was up. I had to run, literally. But, by the time I got to my gate, the results were in. 74 drafts, all there waiting. [music] The next day I started sending. Honestly, I wasn't expecting much. I was wrong. I'm still in the middle of sending these emails. Got 25 left and we've got like seven replies. This is insane. From 54 emails sent, we got 11 replies and booked seven meetings, exceeding our original goal. But, did these founders actually meet with us and hand over their greatest marketing secrets? Subscribe. [music] You won't want to miss this.