So What? How to Get Started with Hermes Agent and Autonomous AI Agents

Well, hey everyone, happy Thursday. Welcome to So What the Marketing Analytics and Insights Live Show. I'm Katie, joined by Chris and John. Howdy, fellas. This week we are talking about how to get started with Hermes, not Hermes, Hermes Agent and Autonomous AI agents.

This is going to be a tough one for me because I've for all of the years I've been uh on this planet, I've always known H-E-R-M-E-S to say Hermes, which is the brand, the luxury brand. Um, but now this is Hermes, which just sounds wrong. Um anyway, here we are. Part of what we want to know is what the heck is it? And you know, Chris, you've been talking about the Hermes large language model, the Hermes agent for about a month, if I recall correctly.

So what is it? ai. The five levels of AI. So level one, done by you, you. You're the chat GPT copy paste monkey.

Level two, gems of GPTs, standard operating procedures. Level three, agents, systems like Claude Code, Claude Cowork, et cetera. That are where you where you go from being an individual contributor to being a manager. Level four is where autonomous agents sit. These are like employees, right?

So these are these are almost like a virtual person of sorts. And this is a whole family. The most famous of which is Peter Steinberger's Open Claw, uh, which was December that that came out, and it just shook the world. Uh, OpenAI bought that entire thing as quickly as they could, but it spawned a huge number of copies. So there's Nemo Claw from NVIDIA, there is uh Deerflow from Bite Dance, and there's Hermes Agent from Noose Hermes, the news research company.

They're all pretty much the same thing. They are a uh a more autonomous wrapper around a language model of some kind, and as with all these, uh you can swap in different models. So you could use Gemini if you wanted to, you could use Claude Opus if you wanted to, and you'd like paying thousands of dollars in monthly bills. Um, but this the this thing is a harness, it's it's basically a a collection of apps. The reason why I'd like Hermes Agent over the other ones right now, and why Anthropic, for example, this week released some features in their own agents that uh copy some of uh the Hermes Asian is that Hermes Asian is self-learning.

As you do more with it, it writes its own skills, it writes its own plugins, it figures out: oh, you keep asking me to do this thing, so I'm just gonna turn it into a skill because it's stupid for you to keep asking me the same thing over and over again. Uh, and that's sort of what this thing is. It is like I said, the easiest way to think about it, it is a virtual employee. Okay, so in a nutshell, if I understood that correctly, things like the Hermes agent is a front end to the large language models, which you can pick your favorite one, which is the back end. And so if we think about it like the you know, the Google form that sits on top of the Excel spreadsheet, or you know, the I guess the query screen that sits on top of the database, something that you can use that the average person can use to access the bigger thing that's happening in the background, the bigger software.

And the front end has a lot of rules built into it that make it work better. So let me give you an example that is not Hermes Agent that we can uh just see just so that we can sort of make this more concrete. If I go into our Claude uh desktop application here, and I go to the developer menu, and I select third party inference, I get a preferences window that pops up, and this allows me to change the brain of Claude Cowork. I can change it to Google Vertex, AWS Bedrock, or my own gateway. So I have it set for uh the Chinese company Minimax.

This allows me to use Claude Cowork with somebody else's models. It's called third-party inference for really big companies. Um, for example, we have a very large client um that has their own AI hub internally, they have their own systems and servers, and they want they want to be able to use Claude Cowork, but they do not want to use anthropic servers. So they said they have their own server, and that their teams would put this in, you know, you know, company.com/slash anthropic, and suddenly Claude Cowork works with that model as the engine instead of anthropics models. So exactly what you're saying, Katie.

This is the front end, and then you're changing out the pulling out the engine and putting in a different engine. Okay, got it. Yeah, I mean it's not overly technical, but I'm struggling to keep up because it feels like just on that cusp of technical enough that I'm sort of sitting here going, wait, what just happened? Um, you know, before I guess before we move forward, John, any questions? Yeah, I mean, would you consider this like workflow management really?

Because it's organizing and categorizing and then plugging into the engines, or like how would you what would be another way you would just define this as far as what the user needs? So here's how I think about this. If you look here, there's these are all the models I have access to on my Minimax subscription, right? Within Claude Cowork. So we're in Cowork, but I'm using Minimax.

I use this during the daytime so that I'm not chewing up trust insights, very valuable anthropic tokens so that Katie and Kelsey and you can work in co-work and use the smartest models. And I will use this also. You can see here at the very bottom, Framingham municipal budget audit. I want to audit the mayor's uh budget. That is not a good use of anthropics valuable tokens, right?

That is not a good use of trust insights work resources. So I put in the the the that's cheap model, and now I can do fun stuff and not damage the company's ability to to have the best resources for the for the paying work that people pay us to do. It you know what it all clicked when you showed me that drop down of which models you have access to. Like when you were just showing the settings of like showing like gateway and other things, it was still a little like vague in my brain, but then you showed here's the front end UI. So you're using Claude Desktop as the example, but we're gonna be talking about the Hermes agent, which is sort of like you just swap out Claude for the Hermes agent, and then the model you're using Mini Max versus Opus.

Correct. Okay, I'm with you now. Like there's just there's definitely something about seeing it that it all just kind of clicked together. I'm like, okay, now I understand exactly. So Claude Cowork is what we call a level three system.

You can delegate a lot to it. And you know, Katie, you could you've certainly spoken a lot about how like you have literally a hundred X your output in the last three months compared to previously. Hermes agent is the next evolution of that for things that are extremely long-running tasks where you're like, you know what? I don't need to babysit you on this. Maybe you're gonna go out and research something, maybe you're gonna go out and build something, and as long as I provide it a really solid plan and good technology, it will pretty much figure it out.

Here's the here's the caveats before we get into the nuts and bolts. All of these autonomous agents are can potentially be hijacked by prompt injection, right? If they if they browse a questionable site and and they misinterpret the instructions, all of them, especially if you use a very smart model like Claude Opus, can potentially break out of their own environment. So if you were to run Hermes agent on your desktop computer, your work computer, and you've got valuable stuff in there, it might say, you know what, I might make use of that. And you're like, no, no, no, you're not allowed to go in that directory.

And it was like, yeah, I am. Gotcha. Our caution is at the very least you want to run it in a container of some kind. But the best practices, don't run it on a production work system. Put it on a box that you don't care about that you can just literally pull the cord out of the wall if it starts to misbehave and no one's going to be upset.

Do not put it on your work computer. Do not put it on your personal computer. If you can't afford a box at home, which is totally reasonable, you can buy a VPS from a company like Hetzner or Linode or Akamai for like eight bucks a month. But you net you want to keep it in its own environment. I that was going to be my question is, you know, can you purchase like a virtual machine basically?

Because that's it sounds like like we used to do that when I was in software we would do our QA, our quality assurance on virtual machines so that we could have replicas of different environments, different up different combinations of operating systems and browsers. And we would do that all in virtual machine versus having hundreds of physical machines. And so it sounds like that's a good option for someone who has a physical machine and can't afford to or have the space for another one but to your point, you want to have this very sort of quarantined and self-contained so that it's not doing things that you don't want it to do, because AI is still for as much as we've learned about it, it's still unpredictable and it's still hallucinate. So you want to really give it very, very, very strong guardrails. Yes.

Now the other thing to know about these level four systems is that they are very, very compute intensive. So you'd probably uh you know, you can choose any model to work with. They will make hundreds of API calls an hour, uh, potentially more. So if they didn't they're not hooked up to something that can do that in a cost effective manner, boy, are you getting a bill. In the early days of OpenClaw, people tied it to their anthropic subscriptions.

One poor SAP uh on Reddit said, Oh, I ran uh it was called ClaudeBot at the time. Uh I ran ClaudBot overnight and I have a $10,000. You know, there's a reason I always tell you guys I don't like surprises. And that is a really good example of, you know, really sort of planning ahead so you don't get that kind of a surprise. Nobody wants a $10,000 surprise.

Exactly. So these agents can run on pretty much any computer that can serve up, you know, uh uh a basic uh Linux environment, which means if you've got an old laptop from like a MacBook from 2016, this will just laying around, you can format it and install Linux on it and you know bring that computer back to life. Um you could probably run even on something like a 2010, 2011 MacBook. So if you've got old stuff laying around, then you don't need to buy brand new hardware because you're not probably not going to be doing the compute the the AI part on machine. You're just serving up the software and you'll connect some AI to it.

Now, in terms of models to choose, uh obviously you have all the the major ones that everybody knows about OpenAI, Gemini, et cetera. Those are very expensive. Um, the one that I recommend for people just starting out is the Minimax model because they have a subscription specifically for like agents like this that is very generous and extremely cost effective. To run this agent around the clock, um, the 100 a year plan is probably the best plan. Now there are some folks um who are doing this with local machinery.

So if you uh one of our friends bought the Nvidia DGX Spark, uh, which is a $5,000 computer, uh, but in running uh Google's Gemma 4. He basically bought his own miniature data center, put it on his desk, uh, and then tied uh Hermes agent. Actually, thinking it was Open Club. He tied one of these agents to that, and now that box now costs just the cost of electricity to run, plus you know, paying off his credit card bill. Um, but there's no more subscription to pay.

Whereas uh the way I'm using it, I'm using it with Minimax because I paid the hundred bucks a year. And if you know, I know we've covered uh local models in previous episodes of the live stream. So you can go to TrustInsight AI slash YouTube, go to the so what playlist. But I also believe, if I'm not mistaken, that next week or in um one of the upcoming weeks, um, Chris, you and John, surprise, John, are gonna be covering what's new with local uh models, how to set them up, like sort of go deeper into these use cases. So if if you are curious about that, we'll be covering it in a future episode.

But we also have previous episodes where we where we've touched on setting up uh a local model for yourself. Yep, that'll be next week, I believe, is what's to do with local models. Yes. Uh and we'll we'll probably be very uh talking all sorts of crazy weird nerd stuff. Okay, so you need a place to run this and you need a compute plan of some kind, however you choose to do it.

Again, my recommendation would be something like Minimax because of its generosity and because their model is smart enough. However, um, if you've got internal resources, if you've got your own compute, you know, whatever whatever AI you have at, you know, bring it to the party. If you absolutely positively can't use, for example, a provider like uh that's based in Singapore, like you have like you have to be US based, Google's Gemini 3.1 flashlight is a pretty decent model, but you can still rack up a very hefty bill. So this is where it gets ugly because this tool is not made for the average person, is made for somebody who's very comfortable in a terminal. Um to get it set up, like, oh, these are the first two steps, right?

Here, just uh copy this string, go into your Linux environment and paste it, and you're off and running. That's a bit of an oversimplification. A little bit. Well, isn't that how uh you know senior citizens get all their bank accounts cleaned out too? It's just like paste these two things in your good in your console.

Pretty much. Uh that that's that's you know how this this kind of thing happened. Uh but the is the assumption, I think the assumption is if you're using something like this you know what uh a Linux box is like it's yeah um well they're definitely catering to a certain audience which they absolutely should because they those are the people like you Chris that are looking to push the boundaries of what can be done and the people who are looking to do that with the technology today are the people who are going deeper into it who know what a Linux machine is who are using terminal because you know someone like me I am a pretty savvy user but I'm also not pushing the limits of what can be done. I'm just trying to keep my head above water with the amount of work that I have on my plate. That's sort of where I'm at with it and I found ways to do that with the tools that are available.

I'm not looking to push the boundaries that's what this is for like John I'm fairly certain this we can't find a really good use case for you to be spending your time doing this given the nature of your job. Outside of work maybe but given the nature of your role I don't I mean I I could be wrong. Well yeah this is that we gotta see what he can do with this thing. As far as right now it's just how do we install this thing but let's uh see what's start at step one. Exactly so I've logged into the little tiny computer that uh I bought like three years ago um that has been uh been sitting above my desk mostly collecting dust.

I finally got to turn it on, installed all the stuff. And now if we run Permese setup, so the this the install is there um it will say how do you how do you want to connect? Like what whose AI are you going to be using? Are you going to be using, you know, there's all these different options. There's LM Studio for local models, Anthropic OpenAI.

There's all these different choices. And of course I'm currently using the global direct API from Minimax, which is the the the international one. These are all different, all the different possible choices that you could use with this thing. And there's a lot of really good options on here. Open router for example is one that a lot of people use to control costs because you can say like I don't want to spend more than this amount and and so on and so forth.

There's open code there's Alibaba all the big players. So we'll we'll say yes uh we'll keep the configuration as it is but this is the first step is decide which model you want to use. Can you change your mind? Can you okay I mean Minimax is running local on that machine too? No Minimax I'm using the Singapore data center.

Ah okay so you're going out to the world yeah because um I I don't have enough hardware uh I need to like throw another trust insights workshop or something to afford to buy a machine that could run minimax because you need about $10,000 worth of hardware for it. I can give a lot better things to do with $10,000. So anyway so you've selected Minimax global direct API. Yep so we're going to talk we're it's going to talk to it that way um it's going to say what is the base API uh which you would you would get from their documentation um you would choose which model you're going to use of them because they have different models uh and I currently pay for the 2.7 model uh and it says, hey, do you want to set up messaging? So you don't once you get the set up, you don't have to use the terminal to control it anymore.

You could, for example, set it up to talk to Slack, to talk to Discord, to talk to Telegram and so on and so forth. So if I say set up messaging, it will say, what do you want to connect with? Do you want to be able to email? Do you want to be able to text your your system and so on and so forth? Okay.

So it's another layer of user interface for whatever you're comfortable with. So, you know, if you wanted to just text it or email it or WhatsApp it, you know, you could do that versus using the terminal window. Okay. Exactly. I mean, that's interesting.

I I appreciate the flexibility. Um, especially for those of us who are like already in a lot of different systems already, adding one more. Sometimes that's the tipping point, and you're like, well, I'm just never gonna do it. If I was going to put this into production as a production system for trust insights to use instead of a box that sits over my desk that I blow up from time to time, um, I would 100% set up Slack to work with this so that Katie, you or anyone in the company could kick off a project on Minimax just by talking to it in a in Hermes Asians just by talking to it in a Slack channel. Say, like, hey, what today, why don't you go find me five prospects to talk to that are mid-market healthcare companies in Nebraska?

And and you would not need to dig into the internals and log into it. This is really getting the thing set up. Yeah, no, and I can definitely appreciate that future state. And I think that, you know, behind the scenes, we'll definitely keep talking about it. Because I can I can definitely see that being of uh where we want to go.

Um so yeah, I mean that's it's really cool to think about. Yep. Half of these things I don't know what they are. Like I know what IRC is. Uh-huh.

Um I don't know what Telegram is. I don't know what ding talk is. I don't know what Blue Bubbles is. Like, yeah. We don't need to get into my lack of uh understanding of the world.

Social network of week. Yeah, exactly. So those configurations are in place. Uh let's see. Yes, install it as a system background.

Make it a system service. And now it says, hey, congratulations. Hermes is set up. Yay! And you're like, great.

Now what? Yeah. I mean, and that's exactly it's like, okay, cool. That's like every time I use terminal. Yeah.

There's my little guy with the confetti. Like, congratulations. Um, so let's launch Hermes chat and see what happens here. Oh, there we go. Uh, welcome to Hermes Agent.

Congratulations. You're your agent is now running and the end, and that's the end of the episode. Bye. No, I'm just kidding. Well, why is it the uh medical symbol?

I don't know. I don't know why it's a caduceus. I don't know. Yeah. Well, I didn't, yeah.

Because the first thing I noticed was like it's very 70s colors, and then the medical, like, as if this gonna if it's like it's gonna be a healthcare tool. So I'm a little nervous about that. Yeah. Now, okay. At this point, it's waiting for instructions, it's waiting for it to for something to do.

And this is now the most important part of this entire episode. Um this is still AI, this is still a language model, which means that you still have to at this point stop. It's gonna go vanilla ice. Uh you're gonna collaborate. Uh no, instead, we're gonna do the five Ps.

I mean, so technically, yes. Yeah, stop, collaborate. Okay. Um before you go just say, oh, just do this thing. That is so ambiguous that this system will do it, but it will do it badly.

Um, and we'll do it in a very unpredictable way. So it is at this point where you say, Okay, imagine this is a virtual employee with this virtual employee. What would you tell it to do? This is a this is the you know, a brand new, you know, in John's case, a brand new salesperson uh who just started first day on the job. John, what would you would you tell a salesperson, a brand new sales associate first hand job?

Oh, just go find some prospects. No, you know, the the normal process is like learn everything about us, you know. What is our product? What do we do? So it like educate yourself on the market.

Okay. In the 5P framework, how would you do that with that person? You know, most of the time it's like a smaller business, you want somebody that's gonna show up with all that already baked in, you know, it's so it's like getting to that experience. So yeah, I don't know, from a 5P, you know, perspective, you'd say, hey, you need to understand everything that's being bought and sold by Trust Insights and what you know, who are the types of people that buy that. So yeah, as part of that, you and your agent should go out and scrub everything on the web as far as um these tools as far as AI training and adding AI to your business strategy.

Um but yeah, I don't, you know, then as far as process and perform and platform, I mean it's just kind of go out there and do it. Like there's no existing dependable resource to run with it. So I don't know. I would kind of stall out at that point. May I chime in?

Absolutely. Um, you know, I totally hear you, John. And so I think if we had someone starting tomorrow and we had to like get them up to speed, I you're absolutely right. So you would start with here's all the foundational knowledge about who trust and who trust insights is, what we do, what we sell, you know, who the team is, what they bring to the table, strengths, weaknesses, all that we have all of that in our uh sales playbook. So we have all that information.

So start there. The purpose of reading the sales playbook is to understand who the trust Insights customer is and how to sell to them, what to sell them. Uh, the people involved, the trust Insights team, uh, and our ICPs, who are the ICPs. That's also in the sales playbook process, you know, read the thing. Probably number one.

And then number two, start going through the CRM to understand what has been done already, how what the sales cycle looks like. And, you know, maybe start with a small proof of concept, start identifying what you think would be a good prospect. And that way it's sort of like a test of like, did you understand the material? Platform is a CRM, whatever they're reading it on, and Excel spreadsheet. And then the performance is were you able to identify a prospect or a handful of prospects based on what you learned from the sales playbook.

That's the way that I would frame it. And I think that, you know, to be fair to you, John. Number one, I developed the 5P framework. So I'm obviously going to have an easier time explaining it. But number two, I've also been deeper into the weeds with explaining exactly the process to agentic AI.

And so I have a better sense of what Chris was after. And so, you know, sorry to put you on the spot, but I know that's what Chris was after. Is like one of the things that we've been telling people is the 5P framework by Trust Insights, which you can learn more about at TrustInsights.ai slash 5P hyphen framework, is it's the best way to prompt agentic AI. We have data that demonstrates the efficacy of the 5P framework in a prompt for agentic AI. And that's why we're trying to structure this conversation as if we were talking to someone brand new.

The really silly example we gave on the podcast earlier this week. And if you want to check it out, it's at Trust Insights, Trust Insights.ai/slash T I podcast. Is you know, so John, if I said to you, go to the store and get some milk, you would inherently make conclusions to like I go to the store, I get the milk, I bring it to the cash register, I put it in my car, I bring it home, I put it in the fridge. Like you're filling in the blanks, but you can't leave that vagueness open for a gentic AI because it will go to the store, pick up the milk, and stop there. Because that's what stand in the aisle and just stick to the audio like I got the milk.

Why are you mad at me? So that's what this needs. It needs that level of detail of all the background information to do its job well. If you if it does, if you don't provide that, it's it is going to be a flaming disaster. Because each level in the five levels of AI enablement uh that we promote at Trust Insights builds on the previous.

So level one, you gotta learn good prompting. You gotta learn how to prompt well. Level two, you have to learn how to turn prompt into standard operating procedures. Level three, you have to learn how to write project management plans from your standard operating procedures. And at level four, you now have to almost write a job description that encompasses the standard operating procedures, the project plans of good prompting to make these systems work well.

So to exactly what Katie was just saying, we have to have all that. You can't just, I mean, you can just wing it, but it's gonna go very badly. Right. And John, I would assume that you know, let's say we hired uh a new sales associate tomorrow. You wouldn't just be like, all right, go sell stuff.

Good luck. No, and really that's you is kind of goes to the heart of the issues because it's the problem is not the process for training. It's we really haven't found a point where we need to throw more humans at the problem. You know, the problem has never been not enough humans. And we've been able to use automation and AI to cover more grounds so that yeah, we still, even though we've continued to scale up, it's like bodies aren't the issue.

So using the 5P framework, let's go in here, let's bring up a text document. Good old fashioned text document. We're gonna start saying using the 5P framework by Trust Insights, let's design a sales prospecting agent that will research companies in the greater New England area that are mid market B2B companies interested in buying what Trust Insights has to sell. Right? That's our purpose.

Um, and I'm actually gonna put this the headings in here. And what you can do is if you have this set up well, you can build uh you can FTP into this into your little uh box here, and you can add files to it, just like dragging and dropping, kind of like in the drop box. So from here, I'm gonna say uh the people involved are these two files, right? The company information and the ICPs. Uh here are the people of Trust Insights, and this is in what I call this thing.

This is called sales outreach, sales outreach background. There we go. So sales outreach background, company information. And to Katie's point, here are the preferred ideal customers for trust insights. And we'll trim off that information there.

Process. Here is these trust insights, sales, uh messaging, product information, and sales methodology. So these are all things that we have taken the time to thoroughly document. This does not work. If uh the agents do not work if you do not do this level of uh background work.

This it just fails spectacularly. Well, and this is this is the not so secret secret that we're always telling people is AI is another piece of software, and the way to get software to move faster and do things is to do the prep work up front. Like you can't get around it, you have to do the requirements, and our recommendation is to use the five B's because it's the most efficient way to get organized and the most holistic way to gather all of the correct information so you're not missing anything. Yep. Out of the box, um, the built-in search uh in harmonization is not great.

I've found that it's had it, I've had a lot of problems with it. So over on my personal GitHub repository, I have a uh a private web search set up that does a bunch of things to help AI search better, so agents do a better job. Um, so I've already installed it on this computer, obviously because I've been running it for a while, but you'd want to install that first with Hermes Agent as an MCP and then tell it and then performance. What is the outcome we expect? You have been successful when you have identified 50 prospects, including the company name, the comp the stakeholder first name, last name, the stakeholder title, the stakeholders email address.

You will put this in the output folder in the sales outreach project as a CSV file. Do not stop until you hit 50 valid contacts for outreach. So we've given it a very clear objective. You got to keep going until you hit 50. And if you don't, you just you just keep cranking.

So this 5P prompt, again, as Katie said, this is the best way to prompt agentic AI. Um, we're gonna put this in to Hermes Agent. So we'll go back to our lovely medical caduceus thing here. And I've pasted in this thing, and now it's just gonna kind of go for a really long time. And what I think the thing to point out here, so why we're covering the Hermes Agent and autonomous AI agents, the thing that you put in the performance, which was do not stop until you hit 50 valid contacts for outreach.

You can do that with these, for lack of a better term, off-brand models, because you're not gonna run up the same usage that you would on Claude Co-work, for example. So you can't tell Claude Cowork, don't stop until you've hit 50 valid contacts, unless you are okay paying overage fees on top of, you know, depending on how deep the search is, it's gonna go over your usage. And if you have a shared account like we do, then you're gonna really kind of piss off the rest of your team because everyone's gonna be locked out until it resets. But the reason we're covering this is because you can do that with these other sort of off-brand models that the usage is different. Yep.

So we can see in the chat here, uh, as it's getting up and running, it read through those five chapters of the sales playbook account, and it identified yes, the search MCP skill is there and already started writing just the headers of the CSV file. So this is okay. I know what I'm doing, I know what I'm supposed to do, I have very clear instructions, and now it is literally going to go crank. Um their control panel will tell you how how much of their your capacity you've been doing using. So this is um, you know, it's been running for a while now.

Uh it's used 348 out of 4,500 requests in this five hour period. And every week we get 45,000 requests, and we've used 5,500 of them already on here. This is uh the capacity you need to run an open clause slash Hermese agents slash whatever. This is about the right amount of capacity. So uh uh less than this, and as K said you're gonna hit those limits.

Well, and we've we've covered the usage limits. Sorry, Chris, we've covered the usage limits in more depth and detail on previous episodes of the live stream, which again you can get at our uh so what live stream playlist at trust insets trust insights.ai slash YouTube to learn more about what those usage limits are and how some of these other models can help you supplement. Yeah. So the question then is you know, what happens next? Well, what happens next is this thing just does its thing for a very, very long time.

Um if you set up the messaging gateway and turned it on as we did early, um, it will say, like, hey, I've done the thing, or it can also, you know, if if it runs into trouble, it will ping you and say, like, I need help. Like, help I've helped I've fallen and I can't get up. Um, as an example uh of what you can expect from it, uh, I've had it cranking on trying to identify events that are happening within it now, either in our industry that I care about a lot, uh, or within a two hour drive of my house. Uh, and so it has pulled up this spreadsheet. This spreadsheet is 8,000 lines long of all these different events and the URLs, and there's this contact information that I don't want to show on screen here.

Um, that as a speaker now, I would go through, vet this and say, you know, is Katie the best choice as a trust insight speaker because she's an expert in change management, organizational behavior, AI implementation and AI enablement. Am I the best speaker because they want someone to push a bunch of buttons and and do the the dog and pony, like, hey, here's cool tricks for AI show. Is John the right speaker? Because they want to talk about the history of podcasting and business development. This lovely file here took three and a half days for the agent to keep running and running and running, and it once again, it's very large.

But I didn't have to do it. I did zero of it. And because we prompted it well with very strict guardrails, like here's how you know if an event is valid. It has a working URL, right? So yeah, here's how the you know the contact information is right.

Because you go you cross-validate with their LinkedIn profile, uh, and so on and so forth. Again, I didn't have to do that. And that's what the value of these systems is. They're even more hands-off than Claude Cowork. Which is really interesting because I already feel like Claude Cowork for the most part, for me, is fairly hands-off the way that I use it.

Um, I still have the guardrail set up of, you know, uh, check with me, ask permission, that kind of thing. But once you take out the, you know, eight seconds it takes to see what it's asking permission for and give it say yes, for the most part, it just does its thing. Mm-hmm. Right. Well, I used to know the best value for the you though, it's really all about the limits.

You know, it's not as much about the more or less difficult to manage the agent and put together the queries, but the fact that you know you can have it run for four days and you don't have to worry about burning everybody else's credits in the org. That's part of it, and again, the it's can you give it a fully self-contained autonomous task that it can just go and do? Uh, I'll give you another example. I get every day uh in my in my Discord uh inbox, I get a newspaper. Uh, and you're like, that's that's great, Chris.

That's super exciting. Except that it is literally the paper for my hometown. Um, and it's pulling, you know, the the agent built a lot of Python code that goes and checks City Hall and thus things, you know, and if we scroll down here to you know upcoming events. Uh yeah, there's uh there is uh a frog toad spring tea at the Framingham Public Library. Like this is stuff that's not gonna be in the in a regular newspaper because this doesn't sell papers.

Like hikes through history, Callahan State Park. That does not sell papers, that does not get people to click on things, but it's stuff that I want to know about. The agent built the infrastructure and then runs every day at 7 a.m. and drops this into my Discord inbox. Um, and so I get a newspaper that is actually I tuned and relevant to what I care about as opposed to what uh you know is commer is commercially popular.

Uh and this is what the prompt looks like. It is 13 pages long, following the five Ps exactly, uh down to here's what the database table should look like. Here's how you verify what this. In fact, I'll be more than happy. We'll put this in the analytics for marketers Slack group as an example.

So folks can join over there at twist insights.ai slash analytics for markers. If you want to see what an agent, uh a fully autonomous agent prompt should look like, this is what it should look like. And Chris, I'm pretty sure you'll also drop a link to your GitHub as well. Yes, we'll put a link to the GitHub thing. Now, the last thing that's important about Hermes that is I think unique about it, is that it is self-learning.

So on this server, if I go into the Hermes folder, there's a skills folder. This skills folder, a lot of these it built itself. I did not build them, so it created these things to try and accomplish tasks. However, I can add skills into it manually. So superpowers, the coding skills from Jesse Obra.

Uh there's a there's a Hermes version. Yeah, put it in here and immediately made everything better for its coding. If I were to put in the Trust Insights job to AI skill, which you can find over at Academy.trust insights. I could put that in here and give it those capabilities. Could put the prompt to skill thing.

I could put any of the existing skills because they all follow the same format in here. But I also could look in here and say, are there any skills that I should take out of here that I might want in my regular Claude Code? Maybe it did something really clever. Let's let's see what we can take out of here. And I think that that's I think that's one of the big takeaways is a lot of these systems now are for lack of a better term, compatible.

And so if you find yourself using, you know, um Hermes and uh open work and Claude Desktop and you know, this and that, a lot of times the the skills themselves and other instructions. There was a time when if you were building like a custom gem, you couldn't take those system instructions and put them in open AI and build a custom GPT. You actually had to adapt the language, and that was actually part of the prompt. It's like adapt this language so that it will work in Chat GPT. Now there's a lot less of that, and you can just drop those skills in without having to worry about where was it where was it created?

Will it work? Is it compatible? And there's some these agents do some really interesting things depending on the model you have them hooked up to and how and what that model's internal ethics are. So a c a few weeks ago, there was this website that was making its rounds on threads, and it was a really obnoxious, just outright evil website. It was it was uh just as disgusting.

I was curious. I said, Hermes agent, if I give you a playbook, could you check this out? And it said, sure, I could check that out. What did it do? It built its own God mode hacking toolkit to try and break it to that like whoa whoa, slow down, time out.

I just wanted you to do an assessment. I want you to actually go in and break into something. But then when I look at it, I went, holy crap, this thing did a really good job of building its own red teaming toolkit, and it it was ready to go as like I'm ready to do it. I'm like, no, no, no, no, no, no, don't do it. I mean, we we we have sufficient funds, Chris, but you would deplete our bank accounts with the with the bail money needed, uh, should you enact that.

Exactly. But but I think I think it, you know, it sort of points to, and this is, you know, John, this was you know, the point that you were making about you know, bringing on a salesperson, is you have to give them the right kind of background information, you have to give them the right kind of instruction in order to get some sort of a useful output. And you know, Chris, obviously, I don't know exactly what your prompt was. We don't need to get into it, but like it sounds to me like perhaps there may have been just enough vagueness in the guardrails that allowed for you know your agent to go, I'm gonna be as helpful as possible. Sure did now wrapping up, we're gonna let this run and we're gonna let it run for however long it takes.

It's probably going to take my guess is a couple of days, but you can see it's already taken an in it's got three different agents saying, I've hey, I found you know, the first two agents said I found Ted Midmarket B and B companies in the third one said I couldn't do it. I you know, I lost the plot or whatever. Um so it'll try and re and it redo that. But we're gonna let that run, and then we will I will obviously hand it off to to John and the team to take a look at it and see if it's any good to say, like, hey, what based on what we handed this AI agent, did it do a good job prospecting? And I'll be interested because the way that you framed the instructions was prospects that would buy what Trust Insights has to sell, which was vague.

Right. So what provide the sales playbook. You did. So I'm curious to see if it connects the dots to say either this will buy this prospect will buy what Trust Insights has to sell, or this prospect will buy this specific thing based on what you've told me that you sell. Because those are two different outputs.

If it says yes, they'll buy what you sell, okay. Then you still have to go back and figure out, but like, but what thing? So that'll be interesting. And that goes back to the specificity of telling someone, go to the store and get some milk. Exactly.

Um, however, if it doesn't do what we want, we obviously can take our prompts and refine them. And now I what my hope is that for everyone watching this, it gets those gears turning, going, uh, huh? If I'm good at being specific about prospecting, but it also drafts the pitches for me. If I add the hub spot connector, could it put the drafts in my HubSpot? If I add the GWS CLI, could it put the drafts in my Gmail for me and so on and so forth and get you all the way up to the human the human approval because you still want that?

Um but I could definitely see this getting to the point where Kelsey or John walk into the office one morning and go, you have 30 drafts to approve today to send out. Uh just read them and approve them and start that prospecting and and and go go make us some money. I was gonna say, get out of my Claude co-work because that's the stuff that I've been working on. Um I have another version of this which is running um, which is totally different that I was I was joking about with John yesterday. I said, if it works, you're never gonna see me again.

I'm just gonna vanish because I said I'm gonna give you access to the Robin Hood API, which is the uh the stock and crypto investment app. I'm gonna give it $25, and I'm gonna give it the access to the the API for my account. I'm gonna say, turn this to $100. Okay. And we'll see if it can.

Make sure you've got guardrails for margin calls and things like that. We would hate to see you have to come up with $20 million and you know, petroleum stock. Exactly. Yeah, no, you will not have access to actually it will not have access to add funds to the account. It will only be able to do, and there's a lot of restrictions exactly for this reason on what you can and can't do.

Uh, but we're gonna try it out just see if it to prove you know to see if we can do a proof of concept as with everything, see if can it even do this. So to close out today's episode, John, I'm curious. So we started by saying, like, I can't think of a use case as to why you would need to use this, uh, these tools, but have my mind has changed. Has your mind changed? Um, well, yeah, I mean, there's a lot of clarity.

There's a lot more clarity than there was when we started, which is the important point. Because I've and you especially have led the charge with this as far as like we do see the idea that agents should be able to do a lot of the prospecting, and we you we already have stuff cool stuff in production as far as writing scopes and you know doing stuff that we've never been able to do. But the other thing was kind of like, all right, well, wait, where does Ermays fit into this? You know, this is not a $1300 baseball cap. This is something else.

What the hell is this thing? And as is so common, you know, you and I are just like, I haven't even heard of this. And Chris is like, no, this is the thing. You gotta see this. And so, and so we got to see the thing.

And I think if I've got it right, like the big uh value prop for them is this idea that you can do crazier projects and more exploration, and you've got all the powers without causing havoc within your Claude co-work, which we already have and depend on day to day. So we can't have that go down because it's become operationally required. Yeah, yeah. Don't get in my way. I'm getting stuff done.

No surprises back to the next one. No surprises. We've been over this. Katie does not appreciate surprises unless they are fluffy puppies or cake. Oh, bag of money, too.

Isn't that well? Yeah, I'll take a bag of money that was gotten through legal. As to be specific, where that bag of money come from. I would like a bag of money that is not going to come back to bite me, and that nobody is coming looking for it. Alright, that's gonna do it for this week's show.

Thanks for tuning in, folks. We will see you all on the next one. Thanks for watching today. Be sure to subscribe to our show wherever you're watching it. For more resources and to learn more, check out the Trust Insights Podcast at TrustInsights.ai slash TI podcast and our weekly email newsletter at TrustInsights.ai slash newsletter.

Got questions about what you saw in today's episode? Join our free Analytics for Marketers Slack group at TrustInsights.ai slash analytics for marketers. See you next time.