Agent Swarms for LinkedIn Pipeline

Julia Nimchinski:

Coming up, we’ve got Maya Voye, founder of Growth Labs, and Maya, I think that you’re gonna showcase the real use case today that everybody’s looking forward to so much, with all of the algorithmic changes of LinkedIn.

Maja Voje:

Oh my god, oh my god. Hey, Yulia, I’m very well, but more concerned about you. It’s your hour number 3 of the event, how are you feeling?

Julia Nimchinski:

Yeah, I’m feeling amazing. We saw a lot of futuristic use cases here, so… Can’t wait to just dive in!

Maja Voje:

Let’s dive in.

Okay, cool. I will share the screen, and hopefully this one will be, like, a little bit easier to do. So right now, do you see my screen?

Julia Nimchinski:

Yep.

Maja Voje:

Marvelous. Cool. I would like to refer to the very first section that we had today, where Amos from Swan AI literally said that we have to earn the right to automate, right?

That we have to talk and, like, literally show tangible ROI. And this is so coherent with the findings of the study that will be published tomorrow, that Kyle Poir and I did.

because AI’s impact is still, like, a mixed bag.

Whereas half of the people participating in our research said that, like, AI is doing very well for them, and they can, like, grasp it perfectly, yet the other half didn’t. So, I think it is really, really, really important that we sometimes consider starting a little bit smaller and focusing on just, like, tangible impact that we can do with either our agents or with swarms of agents.

I like to call them teams of agents. So, this is my very first attempt in which I was trying to solve my very painful problem in business, meaning my LinkedIn pipeline.

I was blessed to have, like, a lot of very interesting posts and experiences on LinkedIn. Right now, 70% of my pipeline literally comes from LinkedIn, but it’s getting tougher and tougher and tougher. And I know that some of the people’s eyes are hurting as they are seeing the final graphics.

It’s only here to demonstrate that by now, I can predictably say, based on even reach of the posts. how many, like, new clients can I get out of them, right? And on my LinkedIn journey, I mean, impressions and organic reach have been and will existing to be, like, a very big issue.

So, last year, performance was 6 million.

From this, I managed to get, like, 150K on my digital funnel. We are doing, like, frequent launches, and we are working with amazing companies, which I’m very proud of, but things have gotten tougher and tougher and tougher. With the renaissance of AI on LinkedIn, and just, like, entry barriers of people using ChatGPT and trying to write, like, Adam Robinson being close to zero, I got a little bit personally upset, but a little bit disappointed as well.

Why? Because I did have, like, what we call a LinkedIn routine, right?

Whoever is posting knows that, that you shouldn’t post and ghost, that you should first do, like, a couple of comments, that you should nurture your posts and be around. But as more and more of these AI comments have been popping up everywhere, I became concerned if I’m the only person left, the only human left on LinkedIn.

And what was, like, an increasingly painful for me was this.

my calendar, right? The majority of content creators in my space spend anywhere from 10 to up to 30 hours a week doing LinkedIn, but everything for me has a very big opportunity cost, because I can literally not spend more than 5 up to 7 hours doing my LinkedIn work. This doesn’t make sense error- wise for me.

So, I really wanted to experiment with AI, how I can bring in agents, that ended up to be a swarm of agents, to literally automate this workflow.

So, these are jobs to be done that I had to do on LinkedIn.

Obviously, you need some sort of strategy, ICP, and a plan. How will you grow? What will you do, right?

I’m very big on positioning USPs and strategic strategic narratives as well. Obviously, you need, like, a pinned-up LinkedIn profile as well, and here, like, the implementation of AI was very basic, so you can use it for writing tests, texts for your profiles and stuff like that, but nothing really fancy. The interesting stuff happened here, right?

So here was the first attempt of us thinking a little bit more systematically.

We created, like, a knowledge base, which serves as, like, this big bucket of all my assets that I have ever created, plus core recordings, and we trained, like. let’s say, let’s call them custom GPTs that help me write.

All my content is still heavy humanly editing, and I’m literally writing substack myself, because I think it just sounds better, no AI is ever able to replace me, just kidding. But really identical opportunities came up in this section.

When it comes to content distribution, to, like, all-bound selling, shouldn’t call it social selling anymore, and just, like, creating very good assets that are aligned with best practices. And here, when it comes to best practices, I can be a complete nerd.

I’m literally making mood boards such as this one. From content creators that I know are relevant for my ICP and are obviously doing better than I am with their 200K followers, I’m trying to learn what are their top, best performing posts. So, what are they doing, and how are they publishing?

So beforehand, I will show you, like, later on, as we discuss the architecture, exactly how this was being done, but this is, like, a major source of intelligence of how I do this work.

And last but not least, I was just, like, super concerned about automations, right?

Because automations on LinkedIn, it’s a nasty word, we shouldn’t be doing it, almost everybody interesting that I know has been temporarily restricted so far.

So yeah, we had to make sure that these type of implementations are safe, so they are not gonna jeopardize my profile.

Now, when I was exploring a couple of stuff out there, like existing solutions, first of all, as somebody who doesn’t know how to code, I was gravitating towards no-code solutions.

I found a couple of ones that are really nice, that can get you up and running, like, in 10 minutes, but the problem was that they became very wasteful in terms of AI token consumption. As well as limiting in terms of, like, the type of workflow that I envisioned, what had to be done. So that was one problem, and then, like, the other pile of solutions just, like, seems too complex to me, right?

For example, whenever people start operating with words like Victor databases, I’m just like, do we need this? No. For simple LinkedIn tasks, like, that we have.

to do a normal database that you can, like, literally self-host for free would totally do the job, so don’t overthink it, don’t overcomplicate it. And especially if you are negotiating for your AI budget, these are the type of implementations that you can do yourself, just like your product manager, or whoever knows how to code with cloth, right?

So I really want it to be as lean as possible.

And what we ended up doing is literally a very strict differentiation between automations, so stuff which is old-school coding, and AI agents.

We didn’t want to just, like, use AI for something that can be much more reliably and cheaper be done with just, like, regular automations that have been out there since forever. So, I created this brief for my team, which is super, super, super tough to manage, but they nailed it.

So, first one, I didn’t want to use any new tools.

Anything that would add complexity to this, like, is a no-no. Second one, zero extra cost. Literally, our, like, ChatGPT consumption in developer account was 46 cents in October.

It was a project that we were doing for a month. And no AI where AI is not need it.

I love AI, but just like in terms of cost predictability for a small pilot, I need to be able to just, like, do stuff automatically, without, like, bartering this with agentic workflows and whatnot. If I don’t have to use the AI agents I want.

That was the principle. And then, I don’t know about you, but I’m a huge fan of dogs.

Do you like dogs, Yulia?

Julia Nimchinski:

Yeah, I’m a doc person.

Maja Voje:

Awesome. So the next section will be very interesting.

I literally see, like, this reinforcement learning as a little bit of this cartoon, right? We have two dogs, Megan and Yuno, and this is, like, one of a very, very, very popular cartoon right now. All the parents will probably know this.

This is Ryder, so he’s managing a bunch of these puppies. that are doing different jobs, and this is exactly how I envisioned our first swarm of agents. And by the way, everything is done in NA10.

I can show you, like, how it’s performing right now. Right now, we have 85% readability, which is not great, not terrible, but nevertheless, like, this is super, super, super, super simple to replicate. You can totally DIY this.

Now, as for the infrastructure and the structure of this solution, this is our first one.

So, the supervisor agents that we are summiting here on Slack is a rider.

Before that gets any way to more complex, this is Ryder. iTech rider on Slack, and Rider summons other agents in order to get the job done. This is how this works.

The next one that are really, really, really interesting are these agents that do the job, right? So, I had 4 jobs to be done on LinkedIn.

The first one is research. The second one is engagement, smart engagement, and I’m still doing human-generated comments. Everything that is coming is coming from me.

Then we have, like, our mini AI SDR chase. and tracker.

This is the analytical tool. Cool. So, as we will be diving further into each and every one, let’s envision this one as, like, before, now the process, how this is done in NA10, and output.

So, what are the type of results that we are seeing here?

So, beforehand, for research, literally, to create a mood board such as this.

I sent these links to profiles to my team, and then they were doing screenshots, and just, like, we were doing a little bit of script, custom script analysis. The project was taking, like, 14 days to complete, sometimes a week if we were in a hurry, but literally, this is, like, what I was doing beforehand. So these are just, like, some of the profiles that I like.

You see Amos here, for example, and Nina, this is very interesting, and I would love to learn what are their best practices. This was my job to be done. So, how we did this later on, in just, like, the Gentic workflow is like this.

So, I’m big on Miro, I’m partnered with Miro, I have everything in my Miro, I call it my digital brains.

And what I want is that whenever I summon it to analyze certain profiles, I want to be able to create a mood board, and I want to see the intelligence.

And I want this notification to be sent to me on Slack.

This is it. So here, it’s just like a development board that we will see later on.

Then we have just, like, a self-hosted server and a couple of custom scripts, so that it generates these screenshots to push in Miro, and then I get notifications. Let’s make it a little bit more fun and see if it will work.

Sometimes it does, sometimes it doesn’t. 85% re-availability. So now, I will summon Ryder to analyze me, Richard King, and Harvey Lee, and Eduardo.

In the last 14 days. Rider will slowly show up, but if it won’t, we can analyze it beforehand, because I did some salmon test it. Interestingly, one of my team members gamified this for me so that I can see the progress.

But we can already move to a previous response that I did when I was testing.

We will see this one as well, but it will take a couple of minutes. So, after this is being processed, I will get two pieces of information back.

The first one is MiroBoard. So, this is a dev account, and this is an extraction of their best performing posts in last, in last 14 days before it, right?

So these are just, like, a couple of visuals, a couple of texts that I can literally take as a design inspiration, or even text inspiration to inform my content.

And the other piece of information that we will get here is detailed analysis.

We haven’t been able to put all this together like we did on the previous mood board, but it’s the same principle. So, we have 3 profiles that we analyze here in the period of 14 days.

We can even, like, play around with the period, and the data will change, so it’s kind of a dynamic. But basically, I would like to learn a couple of things here.

So, what are they posting?

What are their best performing post types?

Then, when are they posting? Hours as well as weekdays is very important. for me, because, organic reach equals new clients, right?

This is what I’m obsessed with. And also, just, like, best performing posts.

And hopefully, these agents here completed by the analysis.

Nope, not yet, still working. So, the end result of this one is literally, as I would perfect it with more profiles, and if you want, I can, like, totally send you this aggregate analysis.

This lasted a couple of hours, and we did design it to bring every piece of information in one place. So that I can, like, literally learn what I should be doing in terms of my content strategy, how many posts should I have, what type of posts, and, like, what is working well from everything that we are experimenting with.

So this would be, like, the first piece of analysis that I would like to run.

Now I’m nervous if it finished the analysis.

Still not. Still working. Something is going on.

But we will return to it later.

It has 15% to finish, and we are there. We will just, like, move to the next one, so that we are straight on time. This is Marshall.

Marshall is engagement agent, and before, how I was literally doing engagement is that I ring a bell on these profiles on LinkedIn, right? And LinkedIn notifications can be quite moody.

Sometimes you get them after 8 hours, sometimes you don’t get them, so that’s a little bit of a nasty setup. Alternatively, I could be doing these saved custom searches, so for example, if I would be interested in GTM, what certain creators like Kyle or even Jonathan posted about GTM, I would be, could be operating from here as well, but this was not very convenient to me, so I wanted to bring this type of notifications to my Slack, so that when I have my posting hour, which is 2PM where I am, I go 15 minutes before, and I just drop a couple of comments, so that I get attention to my comments as well, to my profile, and that I warm up that I’m warming up the algorithm nicely. Oh, and by the way, Sky finished the analysis here, so we will see, like, same data, this development board just showing you that this is working, the same as it did in the setup, and then we have this type of analysis here, which is also very, very, very cute and interesting.

Now, back to Marshall.

Marshall is set up a little bit differently. So, Marshall sends me these notifications of when my observed posts, are posted, right?

So at, 1PM, where I am, this profile’s posted.

This is scheduled to run every 30 minutes, and right now, it’s observing 15 profiles. So then, whenever I’m ready, I can just, like, go and drop a couple of comments.

I do comments manually, because I’m still not convinced by the quality of AI comments, at least under my posts. But yeah, this is really, really, really speeding up my process, so I don’t have to, like, mindlessly scroll throughout LinkedIn to find some posts that I think are worth engaging with.

Now, let’s see how it is built here in the, account.

We will go back to here, and we will just, like, check out Marshall. So, Marshall is actually connected to, this database of new profiles that I have added.

It looks a little bit like a spreadsheet that I have touched or tell you before, and it is scheduled to run every 30 minutes, where it checks the profile. sees if somebody posted, and if it did, it sends me the notifications to my Slack, right? And then, it’s not, like, very intelligent at the moment.

For now, it’s just, like, I have to do 5 comments so that I’m not posting and ghosting. I would love to make it a little bit better, but for now, it does this. Back to the agents.

The next one that we have to observe here is literally my mini SDR Chase. I have no idea if Chase is gonna work or not, but let’s summon Rider once I am doing this type of demo. So now.

I will ask Ryder if he can get me the posts on this event announcement that I did yesterday, right? So yesterday, I published that we have this event, and I would love to know if we have some ICP leads here.

Now, Ryder should pop up any second now to see if it is working.

It is working, and in the meantime, I will just, like, show you for a sec how I did this before, and I will sound like an idiot now, spare with me. So, I was manually checking out the posts, like, who was engaging, and I was just, like, observing their job titles, right?

And whenever I saw somebody interesting there, I was like, hmm, spreadsheet, maybe I should do, like, a little bit of social selling.

Yes, of course, you can do this with integration of clay and heritage, but for just, like, the purpose of my pipeline, so I don’t have, like, a lot of clients, I don’t need much volume, this is a process that would be, like, a little bit overwhelming for me. So, I prefer to query this agent, this, like, Chase agent, whenever I think that this type of topic, this type of post.

could deliver me somebody that is within my ICP bucket. Now, how is this determinated is very, very, very interesting to observe.

Again, we will go and we will check out how chase is being made. Okay, so Chase actually uses, like, a little bit of automations here, for just, like, making sure that we have the right type of people, if they are our ICP fit or not. So this is connected with OpenAI.

We use 04 mini model to use as little credits as possible, but what is super, super, super interesting here is that when I query it, it literally, like, gets… reads all the people who engaged. For now, we have, like, the titles and the countries, very simple integration. We could make it much more complicated, but eventually, it pops up, just like leads that are being processed at the moment.

Luckily, I did it before this event, because live demos are always 5 fun demos, and I know that from this type of post that we did yesterday, we have 20 people that are likely my ICP feed. Yes, of course, I still have to observe this, and for example, this, like, founder from LinkedIn, Ghostwriter, is not something that I would want to social sell to, but there are a couple of very interesting AI founders that I would like to connect with, and I even see a couple of my clients here, which is super nice, because that means that we are within the same ICP bucket. Right?

Of course, in the future, I would love to eliminate that, but for now, we are dealing with this type of setup as just like the first extraction. And, as I was showing you this dummy thingy before, we got another lead engaging with this post.

So now we have 21, yay.

And whenever I’m on the road to just, like, do a little bit of social selling, I would go and just, like, operate with this type of data so that I no longer have to manually browse through everything.

Cool. Now, back to the patrol.

Next, we have something that is still being in the making. So, believe it or not, I was doing my LinkedIn analytics manually.

But it’s great, because it literally forces you to, like, view your data, which is a very cool thing for me.

But nevertheless, it got a little bit boring and repetitive, so I wanted to just, like, create an AI agent.

And I did, actually, the first iteration of this one with RelyApp, and it was, like, rather nice, so this is performing my own post.

But I still wanted to bring everything together, because in the future, I would like to connect these agents even further into a self-learning bundle that would literally take into consideration other people’s posts as well, the ones that Sky is bringing in.

So, that in a nutshell, is the type of setup that I have at the moment.

Again, it’s not perfect. It’s being, like, 85% reliable, but tell you what, it has brought so much joy to me, because literally, when I was doing LinkedIn before, I got just, like, a little bit fatigue, right? It got a little bit annoying.

And this made it fun, and interesting, and engaging again.

So, what I’m telling you during this presentation is literally a very, very, very simple lesson.

before you are just, like, going to these agent swarms and feeling overwhelmed or something like that, just find a little pilot in your business. I don’t care which tool you use, I like N8N a lot. There are, like, other amazing tools on which you can do it, but seriously, can you make your job just a little bit more easy and a little bit more fun using agentic intelligence?

And last but not least.

Please consider where agents are really agents, and where we could literally get away with normal automations so that we are not burning credits like crazy. In this setup, I have saved so much time and frustration. We even managed to close our biggest deal ever, which is a $60 billion company, yay!

I was super happy, because I was just, like, early there, right? This was something that I started as a social media sequence. Is it perfect?

Hell no. Is it something that I will continue building and perfecting? Yes, I definitely want to do this.

So, that in a nutshell, was the first generation of my AI agents that are just, like, producing so much value to my business, and again, what I would like for you to do next is… I’m not selling you any software, I’m literally promoting NA10 that I’m not paid to do.

So, if you want to just, like, get this type of mood boards that we do with agents now, you can scan this code, or we can even share the link in the chat, and you can, like, take it for a spin and see this intelligence by yourself, because this was a very quick demo, and I’m sure that some of you would like to dive deeper into this type of insights.

And yeah, the other one is literally, I’m building this in Pava.

This is something that I really enjoy. This is my favorite projects in Q4, so if you want, you can just, like, continue following our progress on Substack. I think that we are just, like, scratching the surface here.

So, cool!

That was it from my side. We have 5 minutes to do exciting Q&A.

Julia Nimchinski:

Thanks so much, Maya, super impressive, and.

Maja Voje:

It’s not impressive, Yulia, it’s pretty basic. This is something that we did in a month.

Julia Nimchinski:

Wow. Well, Yuri’s… what’s her learning process?

Maja Voje:

of a learning process with building all these agents. Seriously, the first one is that I’m paying a huge tax because I cannot code, right? So, whenever we were just, like, doing this with these DIY solutions, with this plug-and-place, I felt so frustrated and limited, and I really think that right now, you know, those mainstream AIs are getting increasingly great at coding.

And increasingly bad in-text creations.

So, literally, all the code that is written here was written by Mr.

Klotz. And, of course, we had, like, people who are a little bit more technical skills than I am, working part-time on this type of project, but seriously, don’t be afraid of this one.

Don’t be afraid of just, like, asking, like, why are we not open, you know, using this vector database? People on LinkedIn said it is very cool. And then, like, they tell you back because it would cost you a lot of money, and you are like, – I don’t want to do this.

And by now, by just, like, seeing these results, Julia, I really feel, like, more open to invest, right?

Because I kind of feel that this is something that I have experienced value with, that this is something that is literally helping me out, I have hands-on experience now, and now I wouldn’t be hesitant to, like, invest a couple of $1,000 into a proper technology setup, because I have seen it working well for me.

Julia Nimchinski:

That’s great. Maya, just curious, you have a lot of really great customers. Yes.

Maja Voje:

Super successful.

Julia Nimchinski:

Startups, mid-markets, spending the enterprise, so just, curious to hear more customer stories. What are some early innings in terms of results?

Because obviously you’re seeing a lot of results yourself, or at your organization, but what about others?

Maja Voje:

Okay, so one of my customers, Momentum, was already presenting, so we should not get back there. We do, like, content creation together, media. We do have, like, a very conservative client who’s implementing, like, this type of technology, and we are right now chasing a $40 billion company using this type of techniques that we have, presented today, and people already started responding, so we are getting more and more decision-making makers in.

And then, just like for the more scrappy setups, like, seriously, the easiest thing that you can do on LinkedIn is just, like, carriage automations, right?

Again, not paid to promote hairage, but it’s great Because, literally, this sending out DMs and just, like, not waiting and praying that organic reach is gonna be better, it’s not gonna get better, because they just launched Boost Your Own Post function, like Facebook did in 2020, 2012, I think. Right? So we need to make sure that we are shifting our mindset from, like, let’s do content and wait for leads to magically pop up into this all-bound, like, multiple touchpoints, more proactive mixtures, because things are not looking any better, if you ask me.

Julia Nimchinski:

How do you keep up?

Maja Voje:

By seeing my, results, like, my data, by reading, like, a lot of cool substax, this is very interesting. We do have a weekly call with Jonathan, so I’m very blessed to have insights into his special brains as well. But most importantly, Yulia, I read a lot of product management newsletters.

This is where I learned a ton, right? This is where I can, like, literally start developing my taste for tech, so that I’m not like this, there is a new tool, we have tried it, that I can ask in Intelligence questions, so that’s it.

Julia Nimchinski:

Thank you so much, Maya. This was a really fascinating session, and what’s the best way to engage with you?

Maja Voje:

As I said, Substack, where I keep on building in public and hoping to get these agents up and running to do more stuff by each and every month, and definitely LinkedIn. LinkedIn would be nice as well, and yeah, I will just, like, share this link to the board if somebody wants to do, to dive into these insights, to the Slack channel that we have, is that okay?

Julia Nimchinski:

Definitely, yeah.

Maja Voje:

Perfect! Let’s do it!

Thank you so much, good luck with the rest of your event.

Julia Nimchinski:

Thank you so much.