47:00David Ondrej · Tutorial
100 hours of Hermes Agent lessons in 46 minutes
A 47-minute walkthrough of all seven levels of Hermes Agent — from bare VPS to full MCP back end.
May 6thThe argument in one line.
The /goal command makes autonomous agents reliable by inserting a separate judge model that evaluates progress after every turn and halts only when a measurable, verifiable end state is reached -- the only real difference between a useful long-running agent and an infinite loop.
Read if. Skip if.
READ IF YOU ARE…
- You are running Hermes agent or Codex CLI and want jobs that complete autonomously over hours, not just single prompts.
- You want to orchestrate multiple sub-agents without writing orchestration code yourself.
- You are a non-technical builder who wants to use a ChatGPT subscription to power a full agentic setup on a VPS.
- You want to see what defining a clear end outcome looks like in practice with real terminal output.
SKIP IF…
- You are looking for a conceptual overview of AI agents with no live demo.
- You have no interest in running agents on a VPS or a CLI environment.
The full version, fast.
/goal solves the core flaw of earlier autonomous loops by inserting a separate judge model that evaluates progress after every turn and halts when the measurable criteria are met. Hermes's implementation is stronger than Codex's because it supports sub-goals mid-run, survives crashes, and ships as a built-in slash command inside the full Hermes harness. The tutorial covers full setup from VPS provisioning through ChatGPT-subscription auth to two live demos. The one lesson that runs through both: vague goals produce infinite loops; a verifiable end state is what makes /goal actually stop.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:33
01 · Cold open -- what /goal can do
Hook: /goal completes 10-20 hour tasks. Farm produce story. Promise of this video.
01:33 – 02:48
02 · OpenAI drops /goal on April 30 2026
Origin: Ralph Wiggum loop had no judge. OpenAI invented /goal -- give it result, desired outcome, and how to test it.
02:48 – 03:55
03 · Why Hermes /goal is next level
Hermes adds a proper judge model, sub-goals, turn limits. Judge sees goal text + last 4KB of response + system prompt.
03:55 – 05:01
04 · /goal beats a raw Ralph loop -- 5 reasons
Built-in judge, full Hermes harness, crash survival, mid-run interaction, one slash command vs a repo to clone.
05:01 – 06:47
05 · Sponsor + resources link
Free resource bundle. Hostinger VPS setup walkthrough begins.
06:47 – 09:57
06 · VPS provisioning + SSH
Hostinger dashboard, SSH into fresh VPS, Hermes one-liner install. Takes about 2 minutes.
09:57 – 10:20
07 · ChatGPT Codex auth
Quick setup -> OpenAI Codex -> browser URL + 9-char code -> signed in. Hermes now free if you have ChatGPT sub.
10:20 – 14:57
08 · Demo 1 -- AI slide deck via /goal
Create OpenRouter image generation skill, add API key, fetch Midjourney reference, launch /goal to build 5-slide deck. 17 min run.
14:57 – 18:13
09 · Review slide output + SCP to local
SCP .pptx from VPS to MacBook. Open in Google Slides. Midjourney style partially followed, text editable. Harness > weights discussion.
18:13 – 20:00
10 · Demo 2 setup -- CEO/CTO pattern
Hermes has a built-in Codex CLI skill. /goal prompt: Hermes as CEO, spawn two Codex agents -- CTO builds Next.js app, CMO builds B2B campaign.
20:00 – 21:00
11 · Sub-goal mid-run
Demo of /sub-goal to add a 1990s terminal aesthetic without restarting.
21:00 – 23:53
12 · CEO/CTO results + real-world cases
App live on localhost:3000. Twitter use cases: Canon webcam clone in 5 min, 17K record backfill overnight.
23:53 – 25:02
13 · The architecture takeaway
Best setup: Hermes as CEO with multiple Codex sub-agents. Works for code, marketing, research -- as long as goal is measurable.
25:02 – 26:04
14 · CTA -- New Society / Hermes Mastery
New Society community, Hermes Mastery course, 82-use-case diagram, free resources link.
Lines worth screenshotting.
- The only real difference between a useful long-running agent and an infinite loop is a verifiable stopping condition -- not the model, not the prompt length.
- The Hermes judge model reads only three things: the goal text, the last 4KB of agent output, and the system prompt -- cheap, conservative, and unbiased by the executor's own momentum.
- Sub-goals let you steer a running /goal job without killing and restarting it -- new information mid-run becomes a criteria update, not a session reset.
- Session persistence means a /goal run survives a terminal closing or process crash; auto-resume picks up where it left off.
- Using a ChatGPT Plus or Pro subscription as the Hermes inference backend makes the entire agentic stack effectively free for anyone already paying for it.
- Hermes reached #1 on OpenRouter's global token rankings ahead of OpenClaw and doubled usage within three days of that milestone.
- A woman with 300 excess farm tubers told Codex /goal to sell them; it messaged potential customers and sold 208 by morning.
- One developer ran Codex /goal overnight for 3 hours; 16,900 database records were backfilled autonomously with verification passing.
- The CEO/CTO pattern separates concerns correctly: Hermes handles memory, skills, and MCPs; Codex handles sandboxed execution.
- Tool coverage determines ceiling: a Hermes instance loaded with GitHub, Linear, Gmail, and domain-specific MCPs can resolve blockers that would stop a bare installation mid-goal.
- GPT-5.5 is strong at refactors and complex problem-solving; for front-end UI work, Claude Opus 4.7 is the recommended swap.
- Hermes stores API keys in the root .hermes directory so they never appear in the context window sent to OpenAI -- a simple but non-obvious security property.
Define the outcome, not the task.
THE CORE LESSON
Every failed autonomous agent run comes down to the same mistake: giving it a direction instead of a destination.
- A goal without a verifiable stopping condition is just an infinite loop with a better name; tests passing, files committed, and records backfilled are stopping conditions.
- The judge model is separate from the execution agent by design: it reads only the goal text, the last 4KB of output, and the system prompt, keeping it cheap, conservative, and unbiased.
- Sub-goals let you steer a running job without restarting it -- add new criteria mid-run when you get new information instead of killing the session.
- Crash recovery and session persistence mean a multi-hour job survives a terminal closing or process restart; /goal auto-resume picks up where it left off.
- Using an existing ChatGPT subscription as the inference backend makes the entire Hermes stack effectively free for anyone already paying for ChatGPT Plus or Pro.
- The CEO/CTO pattern -- Hermes as persistent orchestrator, Codex CLI sub-agents as executors -- is the highest-leverage architecture: Hermes handles memory, skills, and MCPs; Codex handles sandboxed execution.
- Tool coverage determines ceiling: a Hermes instance loaded with GitHub, Linear, Gmail, and domain-specific MCPs can resolve blockers that would stop a bare installation mid-goal.
Terms worth knowing.
- /goal
- A slash command in Hermes agent and Codex CLI that starts a persistent agentic loop, running until an attached judge model confirms a user-defined measurable outcome has been met.
- Ralph Wiggum loop
- An earlier pattern for autonomous agents where a task is fed into a loop with no stopping judge; it runs until a fixed iteration count, not until the work is done.
- goal_judge
- A separate, lightweight AI model inside Hermes that evaluates after each agent turn whether the goal criteria have been satisfied, returning a boolean done flag and a short rationale.
- sub-goal
- An additional constraint injected into a running /goal session without restarting it; used to steer the agent when new information arrives mid-run.
- CEO/CTO pattern
- An orchestration architecture where Hermes agent acts as the persistent CEO while spawning Codex CLI sub-agents as CTOs to handle discrete execution tasks.
- Codex app-server runtime
- A Hermes toggle that delegates all tool execution to Codex CLI when using OpenAI models, enabling a ChatGPT subscription to replace API credits.
- SCP
- Secure Copy Protocol -- a command-line tool for transferring files between a local machine and a remote server over SSH.
- VPS
- Virtual Private Server -- a rented remote computer with root access, used here to run Hermes agent persistently without tying up a local machine.
Things they pointed at.
25:02productNew Society / Hermes Mastery
Lines you could clip.
18:20
“The clear outcome of the goal needs to be something like tests passing or you have 10 new customers, something measurable. That is the thing.”
17:23
“A model predicts, harness makes it behave.”
17:27
“The best answer often comes from the harness, not the weights.”
23:54
“If you do not do that, slash goal is just a loop.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphorstory
00:00So Hermes agent just released the slash goal feature, and this might be the single most important command of 2026. And that's because with the slash goal feature, you can complete tasks that would otherwise take ten to twenty hours. You can have it autonomously grow your business and actually make you money.
00:15It can refactor an entire code base with hundreds of lines of code in an entirely different programming language, and it can basically do anything that has a clear end objective. And that's why every company, including Codex, Clothcode, and Hermes, is adding their own slash goal feature. So in this video, I'm gonna show you how to set up Hermes agent from scratch, how to activate and use the slash goal feature inside of Hermes agent, how to use Hermes as the CEO with slash goal, and Codex CLI as the CTO also with slash goal.
00:44And finally, how to actually grow your business with this. Because if you don't have real actionable use cases, what's the point? And by the way, this feature is actually insane.
00:52This lady had a farm and she had excess produce. She asked Codex with the slash goal feature to sell it. And Codex went out, messaged some potential customers, and sold her excess produce, actually making her business more money.
01:04Yeah. This is what I mean when I say the slash goal feature might be the biggest AI update of this year so far. Now this all started with the Ralph Wiggum loop earlier this year in January, where basically you give an agent a task, like build for Python course.
01:16It goes to the agent, it saves into a file system, and it runs again. As you can see, this is a very simple loop. But the issue with Ralph loops is that they didn't have a judge to know when to stop, a central control to avoid burning tokens, and a clear goal.
01:28It was just an infinite loop. It wasn't really that serious. But then on 04/30/2026, everything changed when OpenAI released the slash goal feature inside of the Codex CLI.
01:38So OpenAI went back into innovation mode, and they invented this new feature that became useful overnight. I mean, the people that started using it were running it for ten, twenty, sometimes even thirty plus hours building stuff that otherwise would take you days. Because with goal, you can give it the result, the desired outcome, and how to test it.
01:55This is the core difference. And, of course, Hermes wasn't too slow to this. Couple days later after OpenAI released it in Codex, Hermes agent added their own slash goal feature and is arguably even more powerful than the Codex slash goal.
02:07You'll see what I mean by that later in the video. Alright. Now let me explain why the Hermes slash goal is on the next level.
02:13So Hermes has a proper judge that evaluates everything. It allows you to assign sub goals, and it lets you put limitations on how many turns the loop can run for. All of these make it better than the codex slash goal.
02:25So the first thing Hermes did is they introduced an actual judge, a separate AI agent that is completely unbiased from the execution. And for the input, it sees the goal text, the last four kilobytes of the agent response, and the system system prompt. The judge has one simple function to review the progress and evaluate whether the execution agent is headed in the right direction and whether the task is completed.
02:44All of these things make the slash goal feature inside of Hermes much, much more powerful than a simple Rolfe loop. First of all, the slash goal feature gives you a built in judge model.
02:54Second, it lives inside of the full Hermes agent harness. So it gets checkpoints v two, rollback, LSP listing, skills, MCP tools, and everything else that Hermes agent has built in. Number three, the slash goal feature can survive crashes and restarts.
03:06You can just auto resume the goal, and it has persistence across sessions. Number four, it allows you to interact with it mid run. This is very important because when you kick off a goal, it can easily run for six, twelve, eighteen hours, and maybe sometimes you have new information or you wanna steer it into different direction.
03:20In Hermes, you can give it these sub goals and correct it while it's running. Plus, it's a built in slash command versus a whole setup. Right?
03:27With Rolf, you have to clone the repo, configure it, wrap it, blah blah blah. In Hermes, literally comes built in. Right?
03:32So I can literally show you what it looks like. New terminal. Boom.
03:35Hermes. Let me zoom it in, and you just type in slash goal, and you can kick off a new goal.
03:39Now, obviously, I already have Hermes agent set up, so it seems simple, but don't worry. I'm gonna show you everything in this video step by step. You don't have to be a developer.
03:47You don't have to be technical at all. As long as you have a computer and you can speak English, you can absolutely do this. So block out the next forty minutes and pay attention because this might as well change your life.
03:55And by the way, I'm gonna include everything from this video, all the resources, including the image generation prompt, the slash goal prompt from Slidex, the guide for how to use Hermes agent as the CEO, and Codex as the CTO. All of that will be below the video in the second link. You can get it completely for free.
04:08So if you want all the resources from this video, go click the second link below the video and get it now. Alright. The first thing we need is a VPS, and I think we should go with the simplest setup possible, and that is Hostinger.
04:18Hostinger's VPS is not only very affordable, one of the most affordable options out there. They're also so easy to set up. I mean, have a dedicated Hermes agent landing page, which I'm gonna link below, that literally makes it as easy and simple as possible to get your own VPS up and running and install Hermes agent on it.
04:33So so, again, this will be the first link below the video. When you get here, click on deploy, and I will take you to the Hostinger card. Here, the first thing you need to do is select the period.
04:41I recommend going with twenty four months because it'll give you the biggest discount, and you set up these agents once, and they run forever. And trust me, all of us will be running multiple different VPS servers with dozens and dozens of AI agents twenty four months from now. So you better get started now to get unfair advantage over everyone else.
04:56Plus, if you wanna save even more money, go to the right side, click on have a coupon code, and type in David for another 10% off. Hosting your was kind enough to sponsor this video.
05:04So if you use the code David, you'll get additional 10% off of your plan. Then on the left, you can skip daily backups and select the server location, just whatever is closest to you, and then click on continue. In the next step, you'll need to create a Hostinger account and then fill out the payment details to complete the checkout.
05:18Now it usually takes two to three minutes for your VPS to set up. In the meantime, let me show you some of the use cases and some of the latest releases for Hermes agent because the team behind Hermes, these guys, if they can do one thing, they can ship. This is why Hermes agent recently overtook OpenClaw.
05:32Here, as you can see, Hermes agent is now the number one app globally on OpenRouter's token rankings, overtaking even OpenClaw and Kilocode, are massively popular apps. Now Hermes Agent is the most popular.
05:42And just three days later, it's basically a two x. This is how fast Hermes Agent is growing. So if you're not running multiple Hermes Agents, you know, some locally, some on the VPS, some for your business, some personally, you really are falling behind.
05:53Now yesterday, they released this feature which lets Codex be the runtime. Now this release in particular is one that and that's because Hermes can now use Codex CLI as its runtime. So, basically, it's a toggle inside of Hermes that allows it to hand all of the tool execution, shell, file edit, sandbox to Codec CLI anytime it uses OpenAI models like g b d 5.5.
06:13And the reason you should care about this is because it lets you use your ChatGPT subscription instead of API key. So you could already use the OpenAI Codex authentication, which I'm gonna show you how to set that up. So you can use Hermes agent for free since, remember, Hermes agent is fully open source.
06:26It's a free GitHub repository. The cost of it is the AI models. Right?
06:29So if you already have a ChaiGBT subscription, whether that is CheaGPT Plus or CheaGPT Pro, I'm gonna show you how to power not only Hermes agent with that and start using slash goal, but also the codec CLI, which Hermes can control. Okay. So our VPS is finished setting up.
06:43So, actually, we can go to overview, and we can see that, um, no usage yet. It's fresh VPS. But instead of Hostinger panel, if you click on the left on VPS and the one that you just purchased if you're like me, you have, like, 10 plus different VPSs.
06:54So make sure on the one you just purchased, there's two ways to access it. Right? You can click on here and use the integrated terminal by Hostinger, which is probably the fastest way to get started.
07:02But the proper way is using SSH. Sounds super complex, but it's actually very, simple, which is what I'm gonna show you. Now, again, if you don't have your own VPS in 2026, what are you doing?
07:11Go below the video, click the first link, set it up. You can use this for deploying full stack web apps, for running multiple different agents. OpenClaw, Hermes, Agent Zero, Pi.
07:18You can use it for your anything automations. Like, with a single KVM two client from Hostinger, you can do a lot. So go below the video, click the first link, and purchase your first VPS right now.
07:27Alright. So inside of the Hostinger panel, what we need to do is we to copy this command. This allow us to control the VPS through our computer using SSH.
07:35But, again, you don't have to even know what SSH means. All you have to do is open any terminal on your computer and paste this in. Then it asks you, are you sure you want to continue connecting?
07:43Yes. And it asks for the SSH password. Hopefully, you save this during the creation of your VPS.
07:49If not, click on change password and then enter it in. So I'm gonna send in my password. Boom.
07:54And here we are. We are successfully SSHed into our VPS. It took, like, thirty seconds.
07:58Super easy. And here is where we'll install Hermes. So installation is very simple in the Hermes GitHub, which is fully open source.
08:05I'm gonna link it below as well. Everything from this video is in a single link. With the resources, you can get it completely for free.
08:10I'm gonna copy this. Boom. Just paste it into the command.
08:12This is the one liner install command, quick install from Hermes, and we'll install all of the dependencies that Hermes agent needs. Right? Like, back end, the libraries, the packages, everything it needs to run, it's gonna install it on a VPS.
08:23Yeah. This usually takes, like, one or two minutes. So while this is running, let me show you another use case for slash goal, and that is end to end testing in the apps that you're building.
08:31So this guy is developing iOS apps, and he's using slash goal inside of Codex and Hermes to test every single feature in his app autonomously. Just kick it off. It runs for a couple of hours, and it tests everything in an emulator, in the browser, whatever you want.
08:44This is a complete game changer. Alright. Let's check on our Hermes.
08:47Alright. Still installing dependencies. Yeah.
08:48Actually, it's done. Alright. There it is.
08:50It's done. So it took probably, like, two minutes. So next, we're gonna click on quick setup, enter, boom.
08:54We have many different providers we can use. I'm gonna show you OpenAI Colex. That way you can use Hermes agent for free if you already have a ChaiGPT subscription, which hopefully all of you watching already do.
09:03So I'm gonna hit enter. To continue, we need to do two things. We need to open this in our URL, you can just you can just copy this.
09:08Boom. Paste it into a new browser, and log in with your same account used for ChadGBT. There we go.
09:13Now you can either select the team account, if you have ChadGBT teams, and personal account, uh, most of you probably will have one. I'm gonna select personal account because I have the ChadGBT pro plan, which obviously gives me higher limits than the team plan. I'm gonna continue, but it doesn't really matter, to be fair.
09:26Now the second step is the nine character code. So let's switch back to our terminal. As you can see, Hermes gives us this code.
09:31So copy that, paste this in, and click continue. And just like that, you're signed in. So now you can use Hermes using your CHI GBD subscription.
09:38So if you're already paying for it, Hermes agent suddenly became completely free, and it's getting the best OpenAI models. Right? So, obviously, we're gonna go with GPD5.5.
09:47Terminal back end, keep it local. That's fine. We're not gonna set up messaging now.
09:50Skip that, and we are ready to go. So let's type in clear and type in Hermes. We should be able to talk with Hermes agent.
09:56Here we go. Gbt 5.5 loaded. Let's send a test message.
10:00Let's see if we can get a response. Beautiful. Hermes agent is responding, powered by our CherryGPT subscription, so I'm not paying anything extra.
10:08And now we can get started with the slash goal feature. So before we kick off our first slash goal, I'm gonna create a new skill. So I'm gonna say create a new skill for the following.
10:18I'm gonna put in skill. Boom. Paste this long text.
10:22By the way, don't worry. I'm gonna give you this, everything else in the second link below the video.
10:27It's gonna be bundled here together. So you can just copy paste it and give it to your Hermes. And, basically, it will create Hermes agent skill offering.
10:33So it has a built in skill. You we know, can see that Hermes comes with, like, 82 different skills. Right?
10:38Prebuilt. So it already knows how to create new skills. So it uses this one to build it for image generation.
10:43So, basically, this will create full end to end presentations, like, literally professional slide decks. When you show this to people, they'll be shocked. You can give them a reference image, and it will create twenty, twenty five slides in exactly that style, and it looks like PowerPoint.
10:55I had reactions where people thought it's PowerPoint, but, no, it was actually just Hermes generating a presentation using this exact skill. So it named it OpenRouter image generation, and then this will allow us to actually call the skill in the future. So in the meantime, while this is running, we can actually go to openrouter.
11:11We'll just go openrouter.ai, go to the top right, create an account. It's completely free.
11:15Takes, like, twenty seconds. Then we need to go to credits. Make sure to charge up some credits.
11:19You don't need $250. Just do, like, $5. Okay?
11:22Maybe $10. But if you have zero, you will run into errors. Okay?
11:25Then we need to go to the GPT images two model. Right? So it's this.
11:29Already He knows to use this, by the way, but this is the model we're using. It's the latest from OpenAI, the most realistic by far, really, really insane model. And this is the secret of making these amazing presentations.
11:38So what we need is we need to go to preferences, and we need to go to API keys at the top left. I need to create a new key. I'm gonna name the key subscribe.
11:45If you want to see more videos on Hermes agent, make sure to subscribe. It's completely free. And in YouTube studio, the majority of you, over 60%, are not subscribed.
11:53So if you enjoy this video, if you find value in this, please go below and click subscribe. It helps more than you think. I'm gonna put some credit limits like $20 and create.
12:01It's a good practice in case of the API key leaking. Obviously, do not share API keys with anybody. I'm just gonna copy this.
12:07I'm gonna switch back to Hermes, and the skill is created. Okay. So now we can do control c to kill it a few times.
12:12I can do clear. I can say Hermes config set open router API key, and then paste it in.
12:18And it will save it like this. So Hermes space config space set, the name of the variable, and then the value of variable. Right?
12:25This saves it securely inside of the root Hermes directory. So now when you do slash clear and type in Hermes, it can use the key without loading it into the context window. This is very important because it is not sending it to OpenAI.
12:36Right? So remember, we're using g p d 5.5, which means every message we send goes to OpenAI API.
12:41If we were including that key, it's not secure because, you know, people working in OpenAI or if OpenAI gets hacked, your API key could be in there. But Hermes can actually build with it without seeing the API key. So I'm gonna say check this to see if there are any keys in there.
12:55But it's a good to verify that Hermes can see it and can access the key, and then we can actually proceed to the next step, which is choosing a style and generating this full presentation. Beautiful. It contains 12 keys including OpenAir API key.
13:06Amazing. So the next step is choosing the style that we want the full presentation or slide deck to be in. Now in terms of the styles, I think Midjourney is absolutely unbeaten.
13:15Obviously, it's not as good in terms of edits and realism as GPT images too, but in terms of styles, like, look at this. You know? It's amazing.
13:22So, yeah, Midjourney has so many different styles. You can just scroll through the explore page and find something that you like. I can go with this.
13:28So I'm gonna do copy image URL, and I'm gonna switch back into her mess and say, fetch this image from mid journey and save it somewhere on our VPS. Boom.
13:39Give it the URL. Let's see if it can do that or not. Hopefully, it can.
13:42And so it's writing some functions to figure out how to do this, Python commands. There there was an error here. Not good.
13:49But Hermes is very persistent. It it doesn't like to give up quickly. Okay.
13:53There it is. So it fails a couple of times, and I'm gonna say, explain why you failed a few times.
14:00Be concise. Okay. Because Midgenist CDN is protected by Cloudflare.
14:04I see. Okay. Create a skill for this.
14:07Now it can turn these learnings into a skill so that in the future, it doesn't make the same mistakes and doesn't waste tokens on this. This is one of the biggest reasons why Hermes agent is the most popular AI agent in the world and is number one on Open Router even surpassing OpenClaw. We can actually check it here, rankings.
14:22If we scroll all the way down here, you can see the top apps, Hermes agent number one, OpenClaw number two. So, yeah, uh, there's a reason why Hermes agent is the number one because it doesn't give up and it solves things. And it can automatically create skills out of the things it learns.
14:37Right? So now what we can do is we can launch our first slash goal. So I'm gonna do slash goal.
14:41Follow the open air image creation skill. Use the midbrain image. Download it.
14:44Reference. Create a 15 slide presentation. Maybe we can lower that a bit, so I can show you faster.
14:50Let's do five slide. You can easily do 50 slides if you want to. Right?
14:53But, uh, I wanna respect you guys' time. Do not stop working until the full five slide presentation with text on each slide is finished.
15:04Boom. And here we go. We've launched slash goal, and you can see after each turn, a judge model will check if the goal is done.
15:09Her mess keeps working until it is. You pause slash cleared it or the budget is exhausted. You slash goal status slash goal pause slash go resume.
15:16So we can do slash goal status to check the status of the goal. And, obviously, it's reading the OpenAI image generation skill we just created as well as the PowerPoint skill, and it created a plan with four tasks on it. Okay.
15:27So Denture's command, I'm gonna allow for this session. By the way, another pro tip inside of Hermes is you can do slash fast. Toggle on fast mode.
15:34For OpenAI, that means priority processing. And fropic, means, uh, you're gonna be burning credits way faster, but I would definitely recommend you do this. Okay.
15:41So the first slash goal finished running. It took seventeen minutes. The goal was achieved.
15:46We have the five slide presentation. So I'm gonna say show it to me.
15:51Boom. Show it to me. Just that.
15:52It's saved as this file. We'd actually probably give it to the Draxis. Say, like, this is my SSH command.
15:59Tell me how to copy the file from VPS to my machine. There's, like, protocol for that, but, uh, honestly, I forgot it. So I'm just gonna ask Hermes.
16:08Something like SSH, for copying files and uploading files. Let's see what it does, though. It doesn't probably realize that I'm SSHed in.
16:14It pro pro probably try to show it on a VPS. I'm just gonna interrupt it. Okay.
16:17This SCP, that's it. SCP. So this is gonna copy into your local directory, but it should honestly copy into CD documents.
16:26CD documents. Boom. PWD.
16:28This is the path. Change the command so the p p t x goes into on my MacBook.
16:36Well, it'll give me the command. I'll just run the command, you know, in a separate terminal. But at ProTib, you can do command n.
16:41When you're in a terminal to just start a new one, a lot of people kill it for no reason. Let's run this. K.
16:45We need to give the root password. Boom. Paste it in.
16:47There we go. And let's see how fast that is. It took, like, one second.
16:51Alright. So now let me, uh, open the folder in finder. Oh, there it is.
16:54There's the presentation. Let's double click it. We don't have anything to show it.
16:57I'm just gonna drag this to Google Drive and then open it in slides. Open with slides. Let's see it.
17:03Okay. So did it follow the mid journey theme? I mean, you can see the resemblance.
17:08Them. Okay. So it's actual text.
17:10It's actually editable, guys. This is not just a DVD images to generation. A gigantic harness, b drop, model power, .
17:16Okay? A model predicts, harness makes it behave. That's true.
17:19And see, this is editable presentation, guys. The best answer often comes from the harness, not weights. Okay.
17:24Some of these things, like here, it's slightly misaligned, but first shot okay. This slide is cool.
17:29Not bad at all. And this is only the first use case. The second one is even more impressive.
17:33We're gonna use Hermes as the CEO, and we're gonna use Codex in the CLI as the CTO. And Hermes is gonna be orchestrating it in slash goal, and it's gonna be deploying different Codex agents to do development work. So this is completely revolutionary because you can use Hermes from WhatsApp, from Discord, from Telegram to do the work.
17:48And you talk to the agent, and the agent figures out how to deploy these sub agents to best build whatever you want. And, with slash goal, it can run for way longer. I'm talking six, twelve, eighteen, twenty four hours, or even more.
17:59It will run until the end desired result is not finished. So if you take away one thing from this video, it is to do a good job describing the end goal. Right?
18:07The clear outcome of the goal needs to be something like tests passing or you have 10 new customers, something measurable. That is the thing.
18:14Let's just say build a great app. That's not measurable. You need to give it a clear goal.
18:18The judge can make progress. The judge can see, okay. Are we progressing towards that or not?
18:22This is a secret of slash goal. Now if you think this presentation was impressive, wait until you see what's coming because the second use case is gonna be much more powerful. So I'm gonna start Hermes again, fresh session.
18:32And, actually, a lot of people don't realize that Hermes comes with this skill. Again, it comes with, like, 82 prebuilt skills, changes with every version. But look at this.
18:40Available skills, autonomous AI agents, Claude Code Codex. It comes with a prebuilt skill of how to launch and control AI agents in the Codex CLI. So we can leverage that by having Hermes act as the CEO and deploy Codex as the CTO to build any type of app you want.
18:57The beauty of this is that you don't need to have that technical knowledge because Hermes will have it. Hermes will stay in charge of the project. You just give it the main idea.
19:03So I'm gonna demonstrate that live. So I'm gonna do slash goal and paste in my prompt. You are the CEO.
19:08Spawn two Codex sub agents via the autonomous agents Codex skill. CTO, build a modern Next.
19:14Js team weekly reporting app. So you give it, like, the rough overview. What is the clear outcome?
19:19Team members submit weekly reports with wins, blockers, plans, mood. CMO, build a b to b launch campaign, strategy document, free blog post, two call emails. A clear deliverable here.
19:28This is the secret of goal. You cannot have a vague goal. Otherwise, you might as well just launch GBD 5.5 extended.
19:34Right? Extra high. It's great, but it's not gonna run forever.
19:38With goal, it really feels like it can run forever until it hits your goal. This is the most important part. Goal criteria, both repost committed with all required deliverables.
19:47Then the CTO integrates the marketing blog post into the app. So this is the clear deliverable, and now premise is running.
19:52It's gonna do whatever it needs to do to make this happen. So it started by reading the skill codex, planned out six tasks, and then it's, uh, probably needs to authenticate because okay. There it is.
20:01It's installing codex. Right? So it's installing codex CLI by itself, and it's probably gonna gonna copy its credentials to authenticate because I didn't like, remember guys, we didn't install Codec CLI.
20:10We just connected Hermes to our CHI GPT subscription, and now Hermes probably tries to figure out how can I install Codec CLI, which it already did, with my credentials so that I don't have to log in again? This is gonna be interesting if it handles it. So So Hermes is delegating the task to probably do a Codex sub agent, and it launched it with full auto mode.
20:27So not quite yellow mode, but one step below that. And actually, one more thing I wanna show you is the sub goal. So just do slash sub goal, and you can, um, add more criteria.
20:35So maybe, like, say, make the app design like nineteen nineties computer terminal. Boom.
20:42So anytime it's running, you can give sub goals by refining it. So you don't have to interrupt it and launch a new goal. Okay.
20:48So Hermes has finished running. Goal achieved. Response explicitly states.
20:52The work is done. List committed CTO, CVO repose. Okay.
20:55So now what is the simplest way for you to show me these results given the fact I'm SSH ed into your VPS? Be concise.
21:05Just ask Hermes. How can I see the results? Right?
21:07Maybe we can use the SCP again. On VPS CD. Okay.
21:10You can do the on VPS things. Launch these local host servers and give me exact commands to run.
21:19I can also give it back to Hostinger panel. Copy this. This is my SSH command.
21:24Because, obviously, this is a different session than the last one, so it doesn't know. Maybe it can save it into memory. Actually, it's not a bad thing.
21:29Say, by the way, store this into memory as the IP of your VPS. Remember it.
21:36Alright? So now we just do this, I guess, in a new terminal. Boom.
21:41Boom. It needs the root password again. Paste that in, and we are in.
21:45So now we go to locales 3,000, I guess, in my browser. There it is. This is the app, PoolsBoard.
21:50Learn the weaker reporting between PoolsBoard interactive pool.
21:56So I could say, how can I check that the app has all of the qualifications of my original goal? We test something here. It's the CTO and CMO split.
22:05Okay. So the app works. The articles are not written.
22:08That's missing. Make your answer simpler and shorter. I'm not reading all that.
22:11The app slash blog. Let's check the slash blog over here already. Okay.
22:14So this is, way too congested text. Obviously, Codex is not the best at design. You know?
22:20Use Opus 4.7 for design. G p d 5.5 is really great at refactors and solving complex issues, but not front end. So I'm gonna say are both repos on GitHub or just Git?
22:30Okay. So there's just both local Git repos, CO weekly reporting app, CO b to b launch campaign. Yeah.
22:35If we wanted to free up on GitHub, we would need to give Hermes personal access token if it's a private repo. If it's a public repo, then obviously just a link to the GitHub repo, and you can do a PR. So let's look at the execution of slash call because I think that's the most fascinating.
22:48How it started running? The delegation here to the Codex sub agent, you are the CTO. In prod update, I did spawn the two Codex CLI.
22:54K. This was a faster call than usual. Usually, this runs for multiple hours, but this one completed fast.
22:59Maybe the outcome was too easy. Maybe you should have asked for a more difficult app, but you get the point. If you give a clear outcome, slash goal will keep going.
23:08In fact, let me show you a few powerful use cases that people on Twitter have found. Starting off with this guy, Kevin, who used slash goal, this time instead of cloth code, to basically build the mirrorless camera app from Canon, which requires a paid subscription. And he basically built the clone in under five minutes.
23:23And, again, the thing is that I want to keep going until we have the build copy of the Canon Webcam Pro application. Connect my Canon r six. So this is a clear definition of done.
23:30It's knows that it needs to connect the Canon r six, use this PC as a webcam, and needs to look like a Canon webcam pro application. So it will, you know, research what that looks like, the UI, the features of it, and it will keep building until it's done. You need to give it a clear outcome.
23:42If you don't do that, slash goal is just a loop. This guy, Ken Lee, discovered a different use case. He uses slash goal for data backfilling.
23:49So, basically, he ran codecs while he was sleeping. It ran for three hours. 16,000, nearly 17,000 records were backfilled, saving him probably hundreds of hours of manual labor.
23:59And even if you have to code it with AI, still it could be a couple of hours, but a slash goal just to find the end state would you want and it will keep going until it figures it out. So as you can see, the slash goal feature is very, very powerful both inside of Codex and Hermes. And, honestly, the best use case is using Hermes as your CEO and just have multiple Codex CLI sub agents spawn as the CTO.
24:21You can build basically anything as long as you clearly define what you want to happen. And it doesn't have to be in software terms. It could be, I want double my landing page traffic.
24:30I want to contact 20 new leads. Doesn't matter. It could be marketing related, business related, personal related, research related, code related, as long as you define a clear end outcome slash goal will do whatever it can to get it there.
24:44And, obviously, you also want to set up your Hermes in the most powerful way. Because if you set up Hermes with all the tools, all the APIs, all the skills, all the MCPs you can give it, then slash goal can do more things. It will not get stuck.
24:56It will keep going. And to do that, I would highly recommend you join New Society. Because in the New Society classroom, we have a brand new course called Hermes Mastery.
25:05Inside of this course, you get all of the resources to set up Hermes agent, give it its own server, connect it to GitHub, real business use cases, which includes a diagram with over 82 things that you can do today in your business with Hermes Agent. We also have the Hermes Agent toolkit with eight precise use cases that you can implement today of how to actually use Hermes Agent to save time and or to make money or both, obviously.
25:28So if you're serious about AI and you really wanna set up the most powerful personal agent imaginable, make sure to join the new society. Plus, if you join during the month of May, you get a one time bonus, which we're only doing during May, A personalized GitHub audit of any repository you want. But again, this is only available to people who joined the New Society during the month of May.
25:48So if you want us to go over your GitHub repo and get a personalized audit, make sure to join the new society now. It's gonna be linked below the video. And by the way, just to remind you, all the resources from this video are gonna be available in the second link below the video.
25:59Literally, all the skills, prompts, everything I show you is in that second link, so go get it now.
The Hook
The bait, then the rug-pull.
One command. One judge. One question: is the goal done yet? The /goal feature shipped in Hermes agent in May 2026, days after OpenAI dropped it in Codex CLI, and the gap between them is already wide enough to matter.
Named ideas worth stealing.
02:48model
Hermes Judge Architecture
- Goal text
- Last 4KB of agent response
- System prompt
- goal_judge evaluates
- {done: bool, rationale}
The judge is a separate, cheap, conservative model that sees limited context on purpose -- preventing it from being swayed by the executor's own reasoning.
Steal forAny multi-step agentic workflow where you need a reliable completion signal
18:35model
CEO/CTO Multi-Agent Pattern
- Hermes as CEO (memory, skills, MCPs, direction)
- Codex CLI sub-agents as CTOs (sandboxed execution)
- /goal keeps Hermes running until deliverables committed
Separates orchestration from execution so each layer uses the model best suited to its job.
Steal forBuilding apps, running campaigns, or any multi-deliverable project via AI agents
18:20concept
Goal Criteria Formula
A valid /goal prompt contains: (1) the desired outcome in concrete terms, (2) the definition of done (tests passing, files committed, N customers contacted), and (3) any hard constraints (turn budget, sub-agent allocation).
Steal forWriting /goal prompts that terminate reliably instead of running forever
How they asked for the click.
VERBAL ASK
25:02product
“If you are serious about AI and you really wanna set up the most powerful personal agent imaginable, make sure to join the New Society.”
Warm pitch after two full live demos; includes time-limited May bonus (GitHub repo audit). Resources link repeated throughout video.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
