Alex Finn condenses four months and more than a hundred hours with Hermes Agent, the personal AI agent from Nous Research, into nine practical lessons. He argues for running Opus when a task must finish, keeping a second agent as a failover, and stopping the habit of quarantining the agent on its own machine and accounts. The back half covers channel choice across desktop, Telegram, and iMessage, pruning cron jobs to fix a slow agent, wiring a headless device fleet onto Tailscale, and a morning reverse prompt that feeds tasks to a Kanban board. It is an opinionated operator guide with real dollar figures, including his roughly forty dollars a day in Opus spend.
Published Jul 8, 202619:25 video20 min readAdded Jul 11, 2026Open on YouTube →
At a glance
Alex Finn has spent four months and, by his count, more than a hundred hours running Hermes Agent, the self improving personal agent from Nous Research. This video is the compressed field manual: nine lessons on how to turn Hermes from a glorified chatbot into what he calls a full time AI employee. The through line is that the software is only as good as the choices you make around it, the model you feed it, how many copies you run, where you talk to it, and how you keep it fast.
It is opinionated and it is expensive. Finn runs his main agent on Claude Opus and openly reports spending about forty dollars a day, roughly fourteen hundred dollars last month, on that one agent alone. Around that headline sit eight more practical lessons: run a second agent as a failover, stop quarantining the agent on its own machine and accounts, split your day across three channels, prune your cron jobs, wire every device onto Tailscale, interview yourself every morning, and route the resulting work onto a Kanban board. Here is the whole thing, lesson by lesson, with the exact numbers and moves he put on screen.
The premise: an AI employee, not a chatbot
Finn opens with the thesis the rest of the video defends. Hermes Agent, used correctly, gives you your own full time AI employee. Used incorrectly, it is basically just a normal chatbot. The difference is not the software, which is the same for everyone, it is the operator. His promise for the next nineteen minutes is blunt: stick around and you will be a hundred times more productive. Everything that follows is a lever on that gap between chatbot and employee.
Lesson 1: Pick your model like a menu
The single most common question he gets is which model to plug into Hermes. His official answer, if you want best in class, is that nothing beats Opus. He calls it the best agentic model ever made, and the reason is reliability rather than raw intelligence. It is, in his words, the only model where when he hands Hermes a prompt he knows for a stone cold fact the task will finish no matter what. His image for it: give Opus a job and even if it loses a leg halfway through a marathon it crawls to the finish line, where every other model stubs its toe, rolls over, and gives up.
That certainty costs money, because Opus in Hermes runs on API pricing. He has spent fourteen hundred dollars in the last month on Opus credits for his main agent, averaging about forty dollars a day (with one zero dollar day in Cabo when he touched nothing). He says it is worth double that because it runs his business, but he knows most people cannot burn thousands a month, so he lays out the rest of the menu.
If you already pay for a ChatGPT subscription, and most people do, plug that in. OpenAI shipped a real jump at ChatGPT 5.5, which he calls usable with Hermes: not Opus level, not as warm, not as fun to talk to, and it still trips up, but it works. Everything before 5.5, he says flatly, was horrible and not usable. If you want to save even more, GLM 5.2 is a fraction of the price of either and still gets the job done. It feels more robotic, probably because it is distilled from the other two, but it is very good and very cheap, his pick for ultra saving mode.
His recommendation resolves to three answers for three situations. Running a serious business with real revenue: spend the money, go Opus, buy the peace of mind that the task will complete. Already have a ChatGPT subscription: plug it in and spend nothing extra. Squeezing every dollar: GLM 5.2.
The menu
Opus
ChatGPT 5.5 and later
GLM 5.2
Positioning
The filet mignon, best in class
The pick you already pay for
Ultra saving mode
Reliability
finishes the task no matter what
usable, still trips up
gets the job done
Personality
warmest, most fun to talk to
less warm than Opus
most robotic (distilled)
Cost
API pricing, about $40 a day
your existing subscription
a fraction of both
Finn's spend
~$1,400 last month, one agent
flat subscription
the cheapest by far
Use it when
revenue rides on the task completing
you want zero extra spend
you are counting every dollar
Figure 1. The model menu exactly as Finn frames it. The tradeoff is reliability against cost, and his one line rule is spend on Opus when a business depends on the task finishing, otherwise plug in what you already pay for.
Lesson 2: Never run just one agent
This is the lesson he says literally everyone watching should act on: keep at least two Hermes agents running at all times. One is not enough. His main agent is named Hermes (he just likes the name) and it runs on Opus. His second is named GPTme's and runs on ChatGPT. The point of the second, third, or fourth agent is that they watch over each other. Hermes is not perfect and breaks at different times. He finds OpenClaw breaks a bit more and Hermes breaks less, but it still breaks, and when it breaks you need failovers.
His real example: the day before recording, the ChatGPT account behind GPTme's went down, a token expired or something went sideways. If that had been his only agent he would have panicked and had no idea how to fix it. Instead he took a screenshot of the error, handed it to Hermes, his main agent, and said his other agent GPTme's was down and giving errors, could it fix it. Hermes went quiet, came back, said it was done, and GPTme's was back up. That is the whole argument for redundancy: agents that monitor and repair each other automatically.
So plug in any spare accounts you have. If you run Opus and also have a ChatGPT account, add it. If you run ChatGPT and have a Gemini or Google account, add that. He goes further and suggests keeping an OpenClaw agent as a backup too, both so they can confirm each other is up and so you can borrow whatever new feature one platform ships that the other has not.
Setup has two paths. The easy one is to tell an existing agent, in plain language, to build the next one:
Set me up a new Hermes profile. Make it powered by ChatGPT.
Profile is the Hermes term for what other tools call an agent, so a second profile is a second agent. Tell it which model to use, and if you want, name it (he jokes "Name it Harry"), hit enter, and it builds it. The other path is the Hermes dashboard: open it, go to Profiles, click Create in the top right, and walk the flow to pick a model, name it, and launch. Either way you now have a second and third agent watching each other's back.
Figure 2. The failover pattern. Two agents on different models monitor each other. When the backup dropped because its token expired, Finn handed a screenshot of the error to his Opus agent, which diagnosed and repaired it and brought the backup back online.
A quick sponsor note he folds in here: after a hundred hours he says the more agents you build the better you get at building them, and he points to a free eighteen video AI agent building course from HubSpot, the video's sponsor, on the components of an agent (memory, tools, plugging in models, testing and publishing) and the distinctions that saved him headaches early, like automations versus agents and fixed step versus dynamic reasoning, plus wiring agents into tools like n8n, Notion, Make, and Chatbase. You can find that free course through HubSpot Academy.
Lesson 3: Stop quarantining your agent
The third lesson attacks a mistake he sees constantly: people buy a dedicated computer for their agent, give it its own Gmail, its own iMessage account, and separate accounts for every little thing, walling the agent off from their real life. He calls that a massive mistake. You do not need to buy a Mac mini for every agent, and you do not need to hand it its own Google or Apple account. All you are doing is adding needless friction to everything you do.
He understands the fear, and voices it in the audience's words: if I give it my Gmail, will it leak my emails, if I give it iMessage, will it text my ex girlfriend. His answer is that this is not how any of it works. Hermes only does exactly what you tell it to do. Tell it to write a tweet and it will not wander into your photos and leak your nudes. Tell it to fetch the latest AI news and it will not turn around and email your mom. It is an AI, not a sentient actor: you give it a prompt, it does that prompt, full stop. With personal accountability and some thought about what you ask, it will not do anything you did not ask for.
He leans into it. How many people do you personally know who have had a real security incident with Hermes, an agent that went and emailed everyone in their contacts. His claim is that the honest answer for almost everyone is none, and that the security concerns are largely overblown. He preempts the pushback, that a thousand comments will call him reckless, and waves it off: act like an adult, take responsibility, and you do not need twenty accounts bolted onto your agent. Put it on your main computer, on your main accounts, and be a responsible adult.
Lesson 4: Three platforms, three contexts
Where you talk to the agent should match what you are doing. Finn runs Hermes across three surfaces. At his computer he uses Hermes desktop, which he calls the best user experience for an AI agent anywhere. It lets you switch quickly between profiles (essential once you run several agents), see all your cron jobs at a glance, and pin separate sessions rather than dumping everything into one giant thread. You can pop sessions out and talk to several agents at once.
When he is on the go and doing deep work, multi threaded thinking on his phone, he uses the Telegram app. He wishes Hermes had a dedicated mobile app (he notes OpenClaw just shipped one), but for now Telegram is his mobile deep work channel. For quick prompts on the go, the newest addition to his stack is iMessage, which Hermes added out of the box in a recent update. He set up a contact named Hermes, pinned it to the top of his messages, and now fires quick tasks at the agent through the fastest, most always open app he has. The rule of thumb: quick tasks on the go go to iMessage, deep tasks on the go to Telegram, and anything at the desk to Hermes desktop.
Lesson 5: Let Telegram format the answer
Tied to mobile deep work is a feature he wants people to exploit: rich formatting inside Telegram messages. Hermes can now return tables, bold, paragraphs, and structured layout right in the chat. His flagship use is a daily cron job that researches the top AI related stocks, which he tells the agent to return as a table with ratings and details, so every morning he wakes to a clean table of AI stocks to watch. He suggests aiming this at anything that reads better as a grid: tables of stock data, your top performing content, your most important emails. The formatting turns the chat from a wall of text into an organized morning briefing.
Lesson 6: Crons are what make it slow
A frequent complaint he hears is that the agent has slowed down, gone dumb, or is always busy. The number one culprit he finds is cron jobs, the scheduled tasks that run in the background. Like laws and rules, they are easy to create and hard to remember to remove, so they pile up. Every scheduled cron that fires steals time and tokens from the agent.
The fix is housekeeping. In the desktop app, click the cron button to see a clean list of every scheduled job, then pause the ones you no longer need. He does this every so often, and bets you have forgotten half of what you scheduled. Make it a weekly habit: go in, pause anything you are not really using, and you get an instant performance boost plus real savings, because a lot of these jobs burn more tokens than you would guess. The easiest place to prune is the desktop experience.
Lesson 7: Put the whole fleet on Tailscale
For anyone with more than one device, a computer and a phone, a stack of Mac minis, DGX Sparks, whatever, he says install Tailscale. It is free, he is not sponsored, he just loves it. Tailscale builds a private network across all your devices, which lets your Hermes agent move between them. That unlocks local AI models, vibe coding, and simply shuffling documents between machines.
His demo: from the Mac Studio the agent lives on, he asks Hermes to tell him which local models are running on the DGX Spark. Because the Spark is also on Tailscale, Hermes simply SSHes over to it, checks what is running, and reports back. He can then have it download and run a new local model on that box remotely. The payoff is that he does not plug every machine into a monitor. Only the Mac Studio has a display; his five other computers run headless, and the agent operates them for him. This, he says, is where Hermes really becomes an AI employee, working across your other computers on your behalf. If a presentation is stuck on another machine, he tells Hermes to grab it off that computer and send it over, and it moves the file between devices.
There is a bonus for builders: if everyone is on the same Tailscale network and you are vibe coding an app on localhost on your main computer, you can reach that localhost from your phone and test the app on the go. He mentions there is a paid tier but says he has never needed it, he has run on free the whole time.
Figure 3. The fleet on one private mesh. Only the Mac Studio has a display; everything else runs headless. Because every box shares the Tailscale network, the agent SSHes across the mesh to operate the DGX Spark and the other machines, and phones can reach a localhost app for testing on the go.
Lesson 8: Reverse prompt yourself every morning
Hermes is so capable, he says, that you cannot really comprehend everything it can do, which means the hard part is discovering the jobs to give it. His technique is reverse prompting: instead of you prompting the agent, the agent prompts you. Every morning he runs a five minute interview where Hermes asks about his priorities for the day, the tasks he is working on, what is stressing him out, and what is on his plate. It then works out what it can take off his plate, what it can do outright, what it can automate, and what needs his permission, and starts doing the ones it can.
The result, he promises, is two to three new jobs for the agent every single morning that you would not have thought to hand it. He describes the prompt but hands the exact wording out below his video. Reconstructed faithfully from how he describes it, the morning interview reverse prompt does this:
Every morning, before you do anything else, interview me.
Ask me one question at a time:
1. What are my top priorities today?
2. What tasks am I actively working on?
3. What is stressing me out right now?
4. What is on my plate that I wish were off it?
Then, from my answers, tell me:
- Which of these you can do for me right now
- Which you can automate on a schedule
- Which you need my permission before touching
Start on the ones you can do, and list the rest for my approval.
Five minutes in the morning, he argues, saves hours after, because it keeps teaching the agent what it can own for you.
Lesson 9: Send the tasks to the Kanban board
The ninth lesson closes the loop from the morning interview. Once the agent hands you a list of tasks it can do, put them on the Kanban board so the agent can take them over and actually work them. To open it, type hermes-dashboard in your terminal, whether that is Ghostty or the built in terminal, and it pops open the dashboard website. Click Kanban board on the bottom left and you get a board where you add tasks, move them across columns, and assign them to your agent. Feed it the tasks that came out of the morning interview and the agent starts taking care of them. It is his way to stay organized across both the agent's work and his own, and to lay out the day cleanly.
Figure 4. The discovery to execution loop. The reverse prompt interview surfaces jobs you would not have assigned, those jobs land on the Kanban board opened from the terminal, and the agent works them. Run daily, it keeps expanding what the agent owns.
Key takeaways
The model is a menu, not a default. Opus buys certainty that the task finishes, ChatGPT 5.5 and later is the free ride if you already subscribe, and GLM 5.2 is the budget option that still works.
Redundancy is not optional. Run two or more agents on different models so they monitor and repair each other, because Hermes and OpenClaw both break sometimes.
Do not quarantine the agent. A dedicated machine and fresh accounts per agent add friction for a risk Finn argues is overblown for a responsible operator.
Prune your crons weekly. Forgotten scheduled jobs are the top cause of a slow, token hungry agent.
Put every device on Tailscale so the agent can run a headless fleet and you only need one monitor.
Reverse prompt every morning to discover two to three new jobs, then route them onto the Kanban board so the agent actually works them.
Area
Do this
Skip this
Model choice
pay for Opus when revenue rides on the task
assume the cheapest model always finishes
Redundancy
run two or more agents that watch each other
rely on a single agent with no failover
Setup
use your main computer and main accounts
buy a Mac mini and new accounts per agent
Channels
desktop at the desk, Telegram and iMessage on the go
force one channel for everything
Performance
pause unused cron jobs every week
let forgotten crons pile up and drag it down
Devices
join every machine to one Tailscale mesh
wire each box to its own monitor
Discovery
reverse prompt yourself each morning
guess at what the agent can do
Organization
route tasks onto the Kanban board
keep the task list only in your head
Figure 5. The nine lessons as a do and skip ledger. Every row is one lesson from the video, reduced to the move Finn recommends and the mistake he says wastes your time or money.
Chapters
0:00 Intro
0:28 Choosing the right model
3:28 Agent profiles
8:00 Security
10:14 Using the right platform
11:49 Improving performance
14:19 Tailscale
16:45 Reverse prompting
Notable quotes
"Hermes Agent is the most powerful software ever made. When used correctly, you literally have your own full time AI employee." (0:00)
"It is the only model on planet Earth where when I plug it into Hermes Agent, I know for a stone cold fact when I give my prompt to it, it's going to finish the task no matter what." (0:45, on Opus)
"Even if it loses a leg halfway through a marathon race, it is going to crawl its way to the end of the finish line. With every other model, it stubs its toe halfway through the race and then it rolls over and gives up." (1:00)
"You need to have at least two Hermes agents running at all times. Only having one is not enough." (3:28)
"Hermes Agent only does exactly what you tell it to do. If you tell it to write you a tweet, it isn't going to go into your iPhotos and leak all your nudes." (8:40)
"Put it on your main computer, put it on your main accounts, and then be a responsible adult." (9:55)
"This is where your Hermes agent really turns into like your own AI employee, cuz it's basically just on your other computers for you doing whatever work you need to do." (15:20, on Tailscale)
"Take this prompt, run it every single morning. I promise you'll find two to three new jobs for your Hermes agent to do." (18:00, on reverse prompting)
Resources mentioned
Hermes Agent by Nous Research, the self improving personal agent the whole video is about. A plain language walkthrough lives at The Neuron.
Claude Opus, Finn's main agent model, run on API pricing.
ChatGPT from OpenAI, the subscription pick, usable with Hermes from version 5.5.
GLM 5.2, the budget model, a fraction of the price of the others.
Google Gemini, suggested as another account to add for redundancy.
OpenClaw, the rival personal agent, recommended as a backup and now with a mobile app.
Tailscale, the free private mesh network that lets the agent operate a headless fleet.
Telegram, his mobile deep work channel with rich table formatting.
iMessage via Apple Messages, his quick prompt channel, supported out of the box.
Ghostty, the terminal he uses to launch the dashboard with hermes-dashboard.
Claude Code and Fable 5, named at the end as possible topics for future videos.
Alex Finn, the creator, whose channel is all AI content.
Full transcript
Hermes Agent is the most powerful software ever made. When used correctly, you literally have your own full-time AI employee. Here's the thing though, when used incorrectly, it's basically just a normal chatbot. In this video, I will go over every lesson I've learned from using Hermes Agent for hundreds of hours the last 4 months. If you stick with me until the end, I promise you you will be 100 times more productive. Now, let's lock in and get into it.
I have a ton of lessons to go through in this video, but let's start with one of the number one asked questions I get, which is which model should I be using for Hermes? Well, I have an official recommendation for you here. If you want the absolute best-in-class performance, there is no model better than Opus. It is the absolute best agentic model ever made. And now I know what you're going to say, "Wait, Alex, isn't that API pricing?" Yes, yes, it is. I have spent $1,400 the last month on Opus credits for my Hermes Agent alone. Yes, I'm spending a tremendous amount. Yes, on average, I am spending about $40 a day. Yes, I was in Cabo this one day where I didn't spend anything at all, but I am spending a tremendous amount on Opus for Hermes Agent, but for good reason.
It is the only model on planet Earth where when I plug it into Hermes Agent, I know for a stone-cold fact when I give my prompt to it, it's going to finish the task no matter what. The way I like to think about it is this, when you give Opus a task to do in Hermes Agent, even if it loses a leg halfway through a marathon race, it is going to crawl its way to the end of the finish line. With every other model, it stubs its toe halfway through the race and then it rolls over and gives up. Opus, when used in Hermes, no matter what, it finds a way to complete your task. And for me, that is worth a tremendous amount of money. I am willing to spend double what I'm spending now in order to use Opus and Hermes cuz it runs my business so much.
But I get it, not everyone can spend thousands dollars a month on Opus. So, you have other options. If you're already paying for a ChatGPT subscription, which most people are, you can plug that in. ChatGPT 5.5 and later has been very good with Hermes. It's not Opus level. It still trips up. It's still not as warm. It's still not as fun to talk to. It's still not as great at tasks. But it's usable. Everything before 5.5 was not usable. It was quite frankly horrible. Now, if you want to save even more money, GLM 5.2 is also a great model. You can plug it in. It's a fraction the price of both of these, and it'll still get the job done. It's a little bit more robotic than ChatGPT and Opus, probably cuz it's distilled from these other two, but it's still very good, and it is very, very cheap. So, if you're in ultra money-saving mode, I'd recommend GLM 5.2.
So, it's really up to you. You got a menu. Do you go with the best in class, the filet mignon Opus, where you can be sure every task will be done? Do you go with ChatGPT? Do you already have a subscription ChatGPT? That's fine, plug it in. You won't have to spend more money. Or if you want to go in ultra saving mode, GLM. Listen, if you're running a serious online business and you're making revenue, I say spend the money, go with Opus. Your job will get done. The tasks you give it will get done, and that kind of peace of mind is super important. So, my official recommendation's Opus. If you already have an account, just go with ChatGPT. Cost savings, go GLM 5.2.
Second lesson I learned, and this is a massive one that literally 100% of people watching this video should do, you need to have at least two Hermes agents running at all times. Only having one is not enough. Let me walk through this for you. So, Hermes, I named my first Hermes agent Hermes. I just like that name. Hermes is my main agent. It's running off Opus just as we just talked about, but I also have a second agent, GPT-Me's. Why do I have a second one, GPTme's? Obviously, it's running off ChatGPT. The reason why you need at least a second and sometimes third and fourth Hermes agent is they all watch over each other. Hermes isn't perfect. It breaks at different times. I find Open Claw breaks a little bit more. Hermes breaks less, but it still does break. And when it breaks, you need failovers. You need ways to fix it.
So, just yesterday, my account, my ChatGPT account connected to GPTme's, went down for some reason. The token expired or something went wrong. If it was my only Hermes agent, I wouldn't know what the hell to do. I wouldn't know how the hell to fix it. I'd probably panic. I'd probably be like, "Oh, my best friend's gone. I don't know where he went." But when you have multiple Hermes agents, they can all monitor, watch over each other, and fix each other automatically. So, as you can see here, when GPTme's went down, I took a screenshot, gave it to Hermes, my main agent, and said, "My other agent, GPTme's, is down. Giving me errors. Can you fix it?" And literally, as you can see here, didn't have a message again, came back to me, said it's done, and it worked. GPTme's was back up.
You need to have that failover. And so, if you have any extra accounts, using Opus and you have a ChatGPT account, if you're using ChatGPT and you have a Gemini account or Google account, you need to plug them in so that you have those extra Hermes bots available. If you want to take this a step further, I'd even recommend using like an Open Claw as your backup account, too, cuz then you can have them not only watch over each other to make sure they're up, but then you can also take advantage of any features that one or the other comes out with that the other is not using.
Now, how do you set up a second or third Hermes agent the best way? Well, you basically have two options. The easiest way is just go to one of your agents and say, "Set me up a new Hermes profile. Make it powered by ChatGPT." So, profile here, Hermes profile. Profile is the terminology in Hermes. Basically, for other agents, they're called profiles. So, if you set up a second Hermes profile, you're basically setting up a second Hermes agent. So, you just go and say, "Hey, set me up a new Hermes profile." Then, you let them know what model you want to use. If you want to get even more custom, you can say, "Name it Harry." Or whatever you want. You put it in there, you hit enter, it'll build it for you. There's one other way, and that is through the Hermes dashboard. So, if you go into the Hermes dashboard, and then you go over to profiles, you can go in and click create in the top right, and then you can go through this nice user flow to choose a model, give it a name, everything you want. You can put it in there, and that'll build the profile for you as well, and launch it. Now, you have a second and third Hermes agent good to go that will watch over each other. That was my second lesson learned.
Before we get into the third lesson, after 100 hours with Hermes agent, one thing became clear. The more agents you build, the better you get at building them. You start seeing different patterns. My friends over at HubSpot, who have sponsored this video, put together a free AI agent building course just for my audience. You can access with the link down below. It's 18 free videos on how to build your own AI agents. They break AI agents down to every component you need to know, from memory to tools, to plugging in the models. You can see memories and tools here. How to test and publish your different agents. This allows you to go a lot more hands-on on building AI agents. They also talk about a lot of other really important, interesting concepts when it comes to the nitty-gritty of AI agents, like automations versus agents, fixed step versus dynamic reasoning, which would have saved me a ton of headaches early on building these AI agents out, and how to plug your AI agents to the different tools, like N8N, Notion, Make, Chatbase, and a whole lot of other platforms. So, you'll be able to plug your AI agent to every single tool you use on your computer. It's a free, complete course, 18 different videos. The link is down below. Make sure to check it out, and thank you to HubSpot for supporting the channel.
Lesson number three I've learned over the last several months and hundreds of hours using Hermes Agent. The next tip I want to go over is a massive mistake I see a lot of people make, which is this. They buy their own computer for their Hermes Agent. They give it their own Gmail account. They give it their own iMessage account. They give it accounts for every single little thing it does. They completely separate it from everything they do. This is a massive mistake. You do not need to buy your own Mac mini for every agent you have. You do not need to give your agent its own Gmail account, its own Apple account. You are adding needless friction to your entire experience.
And I get why people do, right? They have security concerns. Oh, Alex, if I give my Hermes Agent access to my Gmail, can't it leak all of my emails? Oh, Alex, if I give it access to my iMessage account, won't it text my ex-girlfriend? No, that's not how any of this works. Hermes Agent only does exactly what you tell it to do. If you tell it to write you a tweet, it isn't going to go into your iPhotos and leak all your nudes. If you tell it to get the latest AI news, it isn't going to go then and email your mom. That's not how any of this works. It is an AI. It's not sentient. You give it a prompt, it does the prompt you give it. That's it. So, as long as you have personal accountability and think deeply about what you're telling your agent to do, it's not going to do anything you don't want it to do.
My question to you is this, how many people do you know have had a security incident with Hermes? Truly, tell me. How many people do you personally know have had their Hermes Agent go and then email everyone on their contact list. The answer is probably none. You probably don't know a single person who's ever had a security incident. That's because a lot of the security concerns are largely overblown, and I know what's going to happen now. I'm going to get a thousand comments that say, "Alex, you jerk, you're telling people do this dangerous" Shut up. I don't care. If you have some personal accountability, you act like an adult, you're not going to need to add 20 different accounts for everything your Hermes agent does. You're just slowing down everything you do, adding tremendous amounts of friction. Put it on your main computer, put it on your main accounts, and then be a responsible adult.
The next lesson I learned is what platforms to be using your Hermes agent on. I use three main platforms for Hermes agent. When I am at my computer, I use Hermes desktop. Hermes desktop is excellent. It's the best user experience for an AI agent on planet Earth. You can quickly switch between all your profiles. So, if you're doing multi-agent profiles like I told you to towards the beginning of this video, you can switch to them very easily. You can see all your cron jobs very easily in here. You can pin your different sessions instead of having to having one massive session. You can have separate sessions, pin them, pop them out, talk to multiple Hermes agents at once. It's excellent.
When I'm on the go and I'm doing deep work, I use Telegram. So, I'm still using the Telegram app on my phone, but only for deep work, only if I'm doing multi-threading things on my phone. I hope they come out with a mobile app soon. I know OpenClaw just came out with one. But, if I'm on the go and I'm doing deep work, I'm using Telegram app on my phone. But, other than that, this is a new addition. When I am on the go and I have quick prompts I need to do, I'm using iMessage. iMessage is actually fantastic with Hermes agent. This is a brand new addition as well to Hermes in their latest update. They had out of the box iMessage. I highly recommend setting iMessage up. I added a contact called Hermes to my phone. I pinned it to the top of my iMessages, and now I can message my Hermes agent through iMessage the fastest easiest app that I'm on at all times to get things done quickly. So quick tasks iMessage on the go, deep tasks telegram, and when I'm at my computer Hermes desktop. That's how you need to be using Hermes agent.
The next tip is related to what we just talked about and that is the deep work on the go. There is some amazing new features inside of telegram when you're doing deep work on mobile on the go on your phone and that is formatting inside the messages. You now can have so many nice pieces of formatting in your messages. As you can see here you have tables, you have bold, you have paragraphs, you have different things you can do. You have really nice in-depth formatting now inside of telegram. I'd highly recommend taking advantage of this. Set up cron jobs, ask questions where you ask for generated tables. So I have it do stock research on the top AI related stocks every day. I tell her to put it into a table. I tell her to give it a bunch of ratings and information there and now every morning I wake up to this really nice table with information on AI stocks I should be investing in. Take advantage of this new formatting in Hermes. Set up different cron jobs where it gives you tables of stock information or tables with your top performing content or tables with your most important emails. It makes it a really nice way to use and organize everything going on inside Hermes.
The next lesson learned is around performance. One of the big complaints I get is hey my agent slowing down and it's getting stupid. It's always busy. People have a lot of performance issues with their Hermes agent. What I find the number one culprit to be is cron jobs. A lot of your cron jobs run in the background and much like laws and rules, it's very easy to make new ones. It's very hard to take them away. And so what I find is if you need to improve the performance of your agent, one of the best things you can do is just eliminate and clean up old cron jobs. If you're in the desktop app and you click the cron button right there, you actually see a really nice list of all your cron jobs. And what I like to do is every so often, I go in and I just find which ones I don't really need anymore and I click pause on them. So, you can come in any that you're not really using, pause it, and what happens is when these crons run, it slows down your agent. So, if you have tons of cron jobs going and I bet you forgot about a lot of the ones you scheduled, make sure to just go in every once in a while, once a week, pause any ones you're not really using. You'll instantly get a performance boost on your agent, also save usage and money. A lot of these take up a lot more tokens than you think. So, always clean up your crons and the easiest way to do that is through the desktop app experience.
The next massive tip I have for you, and this is an important one for anyone who has multiple devices. So, if you have a computer and a phone, or if you have Mac minis, or if you have DGX Sparks, any if you have anything more than just one device, you need to install Tailscale. It is completely free. I am not sponsored by Tailscale whatsoever. I just absolutely love this application. Basically, what it does is it creates a private network for all your devices, which allows your Hermes agent to then be able to move across all your devices. This is great if you're running local AI models. This is great if you do vibe coding. This is great if you have a job and you do work and you like to move documents around between all your computers. You can easily go in and just say, "Hey Hermes, go over to this other device and do something." So, if I want to say, "Let me know which local models I have running on the DGX Spark." I can hit enter on that and now my Hermes agent will go from the Mac Studio it's on right now, go over to my DGX Spark, which is also on Tailscale, and be able to tell me what's running on there. So, as you can see, it's SSH-ing over to my other device. It's checking on the models running. And if I wanted to go then go and like run a new local model, download something new on there, run it for me, it can go and do that for me.
It saves me so much time cuz I don't need to plug all my devices into monitors. I just have my Mac Studio plugged into the monitor. My other computers, my five other computers I have, they're not connected to monitors. I just have my Hermes agent go over to those computers and run them for me. This is where your Hermes agent really turns into like your own AI employee cuz it's basically just on your other computers for you doing whatever work you need to do. If I'm ever on the go and I have a presentation on my other computer and it's not on my laptop, I say, "Hey Hermes, go on my other computer, click and drag it over here and send it to me." It'll bring the documents in between computers. You need to have Tailscale. It basically allows your Hermes to run your entire fleet of devices. Has some other benefits, too. If you're vibe coding and you have your app running on localhost on like your main computer, you can have it if everyone's on the same Tailscale network, you can access your localhost from your phone. So, you can test your apps out on the go if you want. So many benefits of Tailscale. It's completely free. There's I think there is a paid version, but I don't even know what's on it. I've just been using free forever. So, make sure to check that out.
My eighth lesson learned is around reverse prompting. Hermes agent is so powerful that we can't even really comprehend everything it can do. But this prompt I'm going to give you down below that reverse prompts Hermes agent helps you figure out how it can help your life the most. Every morning I do a morning interview with my Hermes agent where I basically tell it to ask me a bunch of questions. What my priorities are for that day, what tasks I'm working on, what stresses me out at the moment, what's on my plate. It then takes all those things and then figures out what it can take off my plate for me, what tasks it can do for me, what it can automate, what it needs my permission to do, and starts doing those things for me. Every day I learn more and more about what Hermes agent can do for me, and what it can take off my plate. You need to be doing this morning interview every single morning. It takes about 5 minutes, but it saves you tons of time after because it helps your Hermes agent figure out what it can do for you. I promise you're not getting the most out of Hermes agent. I promise you're going to be automating so much more of your life. You just need to be able to figure out what those things are, and reverse prompting is the way you figure it out. And this prompt that has your agent interview you is the best way to reverse prompt. So, take this prompt, run it every single morning. I promise you'll find two to three new jobs for your Hermes agent to do that will be super helpful for you.
And the ninth tip I got for you, this is an important one. After you do your morning brief, and it gives you a bunch of tasks it can do, put those tasks in your Kanban board for the agent to take over and actually do for you. This is the best way to stay organized on your tasks. All you need to do to see this Hermes Kanban board is go in, type Hermes-dashboard into your terminal, whether using Ghosty or just your built-in terminal on your computer, type Hermes-dashboard. It will pop open this website. You click on Kanban board on the bottom left, and you have this nice Kanban board where you can add in tasks, move them over, assign them to your agent. And so, all the tasks your agent comes up with from your morning interview, you can put it in here, and it'll start taking care of it for you. This is an amazing way to stay organized with the tasks your agent has, with the tasks you have. You put them in here, you're good to go. You can organize your day really, really nicely.
Those are my nine lessons learned. If you learned anything at all, leave a like down below, subscribe, and turn on notifications. All I do is make amazing videos about AI. Let me know down below, what do you want to see next from me? Do you want to see more Claude code videos? Maybe a Fable 5 video or more tips on Hermes agent? Let me know down below. Hope this was helpful. See you in the next video.
