A few months ago, I ran one AI agent. Now, I run a team of them, and between them, they handle most of my business. My inbox, my clients, my content, all run by different bots that talk to each other and hand off work on their own. This is Hermes Bot, and it’s the closest thing to a cheat code I found. So, here’s the plan. First, I’ll break down what Hermes Bot actually is. Then I’ll walk you through building your own team of bots step by step. And at the end, I’ll build a new team of bots right in front of you so you can see how to set one up firsthand. Let’s get started. Okay. In plain terms, a bot is a profile. Those two are the same thing. Hermes has always had those profiles. Every bot you create has its own persona, its own model that you can connect it to, its own tools and access, and its own memory. So, you’ve been able to build those for a while, but now with bot mode, we get the ability to interact with different agents or profiles a lot more easily. The bots can also talk to each other. You can hand one a job. It breaks it up and passes the pieces to others on its own and they can sort it out and talk to each other in one shared room or a group chat. So you stop being the guy running work between these different bots or chats and that is the unlock. It was never the bots, it’s the coordination. Before we build anything, you should know that there are three ways Hermes bots works together. The first is one-on-one where you just talk to a single bot. The second is direct where one bot messages another. And lastly is through a group chat where a few of them work inside of one thread or one group chat with a cap on how many times they can go back and forth so they don’t loop forever and drain you or your tokens. All right, but with that, let me show you how to actually build a team of bots. I always recommend you start with building one bot, just one. Then once you understand how to do that and understand the mechanisms behind each bot, then the sky is your limit. All right, step one to create a bot, you hit new bot and you give it two things, a name and a title. The title is what shows up on the side. The name is what you use to tag it with an at symbol. Then you just describe its job in a sentence or two. And here’s the nice part. From that little description that you give it, Hermes writes the bot’s whole personality for you. You can clone from your default so it inherits your setup, you can pick a model and that’s a bot. That description becomes its soul, its personality and its rules. You can open it up and tighten it whenever you want. So on my sales agent for example, I added lines like never discuss prices unless I approve and always hand the result back to my main Hermes instance. Never straight to me. The tighter you make the job, the better the bot will behave and abide by the job description that you end up giving it. Now the step almost everyone skips and it’s the one that matters the most, especially for saving money. I’m of course talking about tools and permissions. There are three types of tools here. The first is skills. Those are saved workflows and the nice part is they only load when the bot actually needs one. So they barely cost you anything just sitting there. So you don’t need to worry about turning these off. Then there are tools. Those are live actions your boss can perform like reading a file, searching the web, sending a message and those load into every message. And then there are MCPs. Those are external tools that plug your Hermes onto external tools like Gmail, Notion or GitHub, but those also cost you a ton of money if all of them are turned on. Every bot starts with every tool in MCP switched on, and you do not want that, or rather every tool in MCP that your main bot is uh connected and turned on to. The more tools your bot has enabled, the more context it will use for every run it has. Some tools have upwards of 50,000 tokens that load into every single message you send. So if a bot does not need access to a certain tool or MCP, then disable it. You give each bot only the tools its one job actually needs, and you turn the rest off. This will make your bots faster, cheaper, and more efficient. This is also the step that takes a bot from a decent bot to something that is genuinely good because the MCPs you set for each bot are extremely important. Think of it as handing one bot a specialist power that the others just don’t have. For my sales agent, for example, I wire it into our sales CRM. This gives it access to edit, to search, uh, and to sort through all of our leads that no other agent has or for that matter needs. It’s pulling from a tool built for that exact job. You add it once in that bot’s capabilities, and from then on, it’s part of its kit. One bot, one upgrade, and your entire bot team’s output just simply gets better. All right, while we’re in settings, there is another setting you can change to save money, and it’s setting each bot’s model. Your orchestrator, aka your your main bot, aka the bot making the calls. You want to put that on a strong model. Think of putting one of your strongest and most capable models here because it will need to be able to make decisions and think through complex reasoning problems. The grunt work bots or the worker bots on the other hand, drop them to a cheaper model or run them on a local one on your own machine for free. You will essentially get the same exact result on the simple stuff for a fraction of the cost. And that’s mostly what keeps a team like this cheap instead of pricey. You’re only paying for the smart model where it actually matters. All right, two more quick final settings to adjust here. For memory, let only your orchestrator write to your shared memory if you have a shared memory provider like Honcho or Memo or Super Memory. And for every other bot, switch that to built-in memory only. We do this because our main orchestrator already saves the important information to the shared memory. So there’s no need to double save it and waste context for no reason. In settings, you can also choose to bump how many bots stay warm from a three to a five. This will make it so switching between bots is instant and they won’t need to reboot every time you open more than three bot chats. And finally, timeout time is by default set to 10 minutes. I personally recommend you set it to 20 or 30 minutes depending on how big your workflow or how much time your workflow typically takes. This will give your bots a longer idle timeout so a long job doesn’t quit or error on you halfway through. Finally, this is the last step to make an ultimate Hermes bot team and it is defining anything you never want an agent doing without your input. For me, those are three things. spending money, sending something to a client, and deleting stuff. Those are the things I always personally want a hand on. Everything else, it can just run itself without my approval. So, before you finish building your team, make your own list. Ask yourself, what could a bot do that I’d actually regret? Those right here, your answers, those are your gates. It will be different for everyone, so be honest about yours. Once you have your gates, you have three places to set them. The first and the strongest one is what tools a bot even has. If a bot doesn’t have the tool to send an email, it can’t send an email, no matter what it decides in the moment. So, the cleanest gate is just never handing a bot the keys to something it should never touch. The second is Hermes own approval setting. You can put it on smart and it’ll auto run the safe stuff and block the genuinely dangerous stuff and stop and ask you on anything in between. When it asks, you get four choices. Allow it once, allow it for the whole session, always allow it or deny. And the third layer is inside the soul.md file itself. You can write the rule straight into the bot’s personality. You can say something like never send anything to a client without my approval. Just know this one is a softer layer, meaning the model will try to honor it, but it’s not a hard wall like the other two that we just covered. So, use the soul to reinforce, but never as your only lock. Setting those guard rails is exactly what lets me hand the team my real business and be able to sleep at night. They run free, but the stuff I’d regret is gated behind a yes from me. Okay, enough setup. Let me put everything we just learned together and apply that. I’m going to build a content team that consists of three bots. My main Hermes agent runs point as the orchestrator or the operator. Then we’ll need a researcher whose only job is going to be to dig into what’s working right now in my space and a content creator that takes what the researcher finds and turns it into actual posts or content. So, let’s start by asking our main Hermes instance to do just that. Hey, Hermes, I’d like for you to help me build a team that will post and manage content creation on an Instagram page for me. Uh, you can ask me as many questions as you would like until you reach clarity. Ultimately, I’m thinking we could have two Hermes bot employees. The first one, a researcher that monitors what is working in that specific niche. And the second one, a content creator that will go ahead take that information and create content posts, graphics in the approved styles that we end up choosing and the direction that we ended up going for. Ultimately, I want you to be the one telling me um or relaying the information back to me. I don’t want to be speaking to the other bots. I want you to be my main point of contact. Can you help me out with that? And we’ll send this out. All right. So, I’ve given it the context. It’s asked me a bunch of questions and it’s currently starting to delegate and plan it out. Got it. I’ve locked it in. It has the brand. Please send me the exact Instagram URL. And it asked me a few questions. I’m going to go through this onboarding process, as I like to call it, and come back once we’re done. Would you look at that? It literally just ran through a set of tasks. Oh, researcher hit an error. That’s a very nice start. It ran through a set of tasks and it created them. I’m It even created the group chat itself. Uh to which I’m I’m just saying hi to my new team. Never mind. They did end up replying. I waited like 5 minutes and none of them replied. So, I went to check that my Hermes is still live. It was. I’m not sure what happened, but there we go. Hey everyone, what are we cooking today? Let’s go. Just as simple as that. Turns out what happened was our provider timed out, which is GPT. Uh we reached our uh session limit. Uh so I had to quickly reset that. Now we’re back to being up and running. All right, so it finished setting them up completely. It created the group chat. Everything is done. The one final check that I want to make before we get started with this workflow is I want to inspect the soul that it gave it mission. Very nice. It gave it a mission, a reporting line responsibilities. This is a very well-defined soul. Much better than what I could have done. Uh, and it’s on our subscription, which is good. It’s plugged into a cheap model, so it doesn’t end up costing us a fortune. Now, for skills, we don’t really care too much about that. Remember, what we want to do is turn off any tools or MCPs that that bot does not need. This is the creator bot. The creator bot will need access to Instagram. So, I’m going to connected to Instagram and disconnect it from anything it doesn’t need. Then, I’m going to go to the researcher and do the same thing. Uh, add and remove whatever tools, MCPS, and plugins it needs or doesn’t need. and then I’ll be back once that is done. All right, done with that. I’m going to ask them team let’s get started. Now what you can do in that group chat say you want to reach out to the researcher specifically then you type at researcher and just like that you can tag it so that it knows this is for them specifically. I can tag the creator. I can tag all everyone or I can tag the orchestrator operator specifically. And there you go. Just like that, it’s following our instructions, which is Hermy the orchestrator/operator is the one I’m talking to. He’s coordinating everything. And he said, I’ll coordinate and quality check the final package at researcher. So, it’s delegating to the researcher. Surface today’s strongest evidencebacked psychology angles. and at creator it’s delegating to the creator turn the best one into a sharp practical concept once the research land I realized I didn’t brief you on this but I used to run a psychology page called daily psych on Instagram a few years back and the job that I gave them was to brainstorm and come up with a plan of content angles that they can create to revive the page and just like that the researcher is doing his job, the creator, strong handoff. They’re talking to each other and you could just sit back, watch, come back in a few minutes and see what the back and forth was like. And this is great because they understood the assignment. They came up. So, this is the the chat. They had a back and forth. The researcher came up with the research, handed it off to the creator. The creator conceptualized how they can make that into content. Then they sent it to Hermy. Hermy, here’s the finished carousel draft. And Hermy, relate this back to us with a caption, with a style. All I asked was great. Now you can create the post and run them by me here. A nice thing you can do over here is if you have multiple workflows running, you can create groups. So I’m going to call this daily psych. And inside that group, we’re going to move, there you go, all of the bots that are meant to be in this group. That way we can stay organized around all of our different groups. So say you have multiple content pages, you have multiple, I don’t know, businesses, workflows, uh, levers that you’re pulling, you can organize them that way so that they’re much easier to navigate through. Sorry, but this back and forth that you can watch your bots just go through is insane. It’s crazy how fast this tech and innovation is moving. Essentially with that, this is what they ended up creating, which I absolutely love. This is the first slide, the second one. So, they created the carousel, sign two, sign three, try this tonight, and a call to action. Save this for the next hard day. And it’s insane to watch the back and forth they’ve had. Once the creator finished the content, he sent it to Hermy, said, “Hey, I’m done.” Hermy then tagged the researcher and said, “Hey, like check it or doublech checkck it uh for anything that might be off or not evidence-based.” The researcher checked it and said, “Hey, no research objections in this copy packaged as drafted and Hermy came back and said, “Hey, I’m done.” That is insane. Here’s the part that makes this entire process compound. After a job wraps, we can tell the team to save what it just learned. So the next time I ask for the same kind of thing, it already knows my format, my style, the sources that I trust, and it leans on me way less. At the end of the day, you’re not just running a team. You are training one. This is how you eventually get to it, asking you less and less questions with time and need less and less of your input. So here is my advice for you. If you have a one-off job, don’t even bother with a bot team. One good agent can do that job. I mean, you don’t need a team of agent to do one simple task. For example, like manage an email. One really good bot is more than enough sometimes. Don’t go build a whole team just to answer one email is what I’m trying to say. But the things you do every single week, managing your content or your business or your weekly workflows, that’s where building a team of AI bots really shines. a handful of specialists, each locked to their own job, checking each other’s work and running while you sleep. That’s the whole point. The best part is that Hermes is free and it’s open source. So, the only thing this costs you is setting it up, aka your