Learn 95% of Hermes Agent in 31 Minutes

0:00Hermes Agent is one of the most powerful

0:03AI tools in the world right now, but it

0:06has one of the worst beginner problems.

0:09You install it, open it, and then

0:11immediately realize you have no idea

0:14what you're actually supposed to do with

0:16it. Because Hermes does not behave like

0:19a normal chatbot. It can live on your

0:22desktop, it can run through Telegram, it

0:25can remember how you work, create

0:27skills, use tools, schedule jobs, spin

0:30up different sub-agents. It can run

0:32different specialist profiles and

0:34basically become this AI operating layer

0:37around your life and business. That

0:41sounds amazing until you are staring at

0:44a blank chat box thinking, "Okay, what

0:47now?" So, in this video, I'm going to

0:50teach you 95% of what actually matters

0:53in Hermes Agent. I'm going to give you

0:56the mental model, the setup, the model

0:59choices, the memory system, the skills

1:02system, the tools, the scheduled jobs,

1:05the sub-agents, the profiles, and the

1:08real workflows people actually install

1:10Hermes for. By the end of this video,

1:13Hermes should stop feeling like a

1:15confusing AI toy and start feeling like

1:17a real assistant that can operate your

1:20entire life. Let's get started. The

1:22easiest way to understand Hermes is

1:26this. ChatGPT is a place you go for

1:29answers. Claude, Code, and Codex are

1:31agents usually point at a specific

1:33project. Hermes is trying to become the

1:36layer that connects AI to your actual

1:40work. That means instead of opening a

1:42new chat every single time you need

1:44something, Hermes can sit across your

1:47tools. It can remember your preferences,

1:49use your files, run scheduled work, and

1:52talk to you from the places you already

1:55live. For me, that means Telegram,

1:58Notion, my daily briefs, my memory wiki,

2:01and a bunch of internal business systems

2:04that I use every day. The important

2:06thing is that Hermes is not valuable

2:08because it gives slightly better

2:10answers. Valuable because it can do

2:12repeatable work and improve every single

2:15time. That is the mental shift. If you

2:17treat Hermes like a chatbot, you will

2:20ask it random questions and be

2:22disappointed. If you treat Hermes like

2:24an operating layer, you start giving it

2:26jobs. Research this topic for me every

2:29morning. Turn this process into a skill.

2:31Draft responses every morning, but don't

2:34send anything until I approve. That is

2:36where Hermes starts to click. It is

2:38conversation plus memory plus tools plus

2:42scheduling plus workflows. That

2:45combination is the product. The next

2:48thing that confuses people is where

2:50Hermes actually lives. Because there is

2:53Hermes in the terminal. There is Hermes

2:56desktop. There is the dashboard. There's

2:58Telegram. There are other messaging

3:00platforms. And if you're new, that can

3:02make it feel like there are five

3:04different products. But there are not.

3:07They are different surfaces for the same

3:10agent. Here's the simple way to think

3:13about it. Hermes desktop is the control

3:15room. This is where I would start if

3:18you're new. You can see your sessions,

3:21switch models, manage profiles over

3:24here, and use Hermes in a way that feels

3:27closer to a normal desktop app. Telegram

3:30is the daily assistant surface. Some

3:32people might even choose for this to be

3:35an iMessage and Discord and WhatsApp or

3:38wherever. This is where Hermes starts

3:40feeling different from every other

3:42agent. You can send it a voice note

3:44while walking. You can message it from

3:46your phone. You can have scheduled jobs

3:49show up as a

3:51Then the CLI and the dashboard are the

3:54power user layer. This is where you

3:56configure, inspect, debug, run more

3:59advanced workflows. This is where you

4:01can create your own mission control,

4:03just like I have over here. You can use

4:05commands like Hermes setup, Hermes

4:08model, Hermes doctor, so on and so

4:10forth. The mistake is thinking you need

4:13to master all of this on day one. You do

4:16not. Start with desktop, connect one

4:19model, connect Telegram or whatever

4:21messaging [music] platform you naturally

4:23already use if you want the phone

4:26assistant experience, and run one

4:28workflow through that. Once that works,

4:31then you can care about the rest. Here

4:35is how to set up Hermes very quickly.

4:38You can open Hermes's official website

4:42and just press on download. I'm on a

4:44Mac, so it shows me download for Mac OS.

4:47For you, if you're on Windows, it should

4:48show you download for Windows, or you

4:51can even run the command locally inside

4:54your terminal. If you have a dedicated

4:56spare local machine or device, you can

4:58go to the Hermes agent website, click

5:01install there, and download whichever

5:03version is compatible with your

5:04operating system. Open it and run its

5:07installation. Really, it's as simple as

5:10that. If for any reason something

5:12doesn't work, you can open a terminal

5:14window on your machine and type in bash

5:17Hermes doctor. This tells you what is

5:20working, what is missing, and whether

5:23your provider, tools, gateway, and

5:25environments are healthy. If you don't

5:28have a dedicated machine, then you can

5:30install Hermes on what we call a cloud

5:33server or a VPS. There are a ton of

5:36Hermes VPS providers where you can

5:38choose a plan, and they have a one-click

5:41install Hermes through them. Then, the

5:44next big thing is you'd need to choose

5:47one main model that Hermes would use. Do

5:49not overthink this. At the start, pick

5:51one strong model that can actually use

5:53tools well. We'll dive deeper into model

5:56selection in just a minute, so we can,

5:58you know, talk more in depth about that.

6:01Then, I would open Hermes Desktop and

6:03make sure that I can start a session.

6:05Then, I would connect Telegram or

6:07whatever messaging platform you already

6:08naturally use if I want Hermes on my

6:11phone. Then, I would run one stupidly

6:15simple test. Something like, "Create a

6:17short daily briefing template for me.

6:20Ask me three questions about what I want

6:22included before you write it." That

6:24sounds basic, but it tests the most

6:26important thing. Can Hermes understand

6:28the task, ask for missing context,

6:31create something useful, and preserve

6:33the decision trail? The beginner mistake

6:35is spending 2 hours on configuration

6:38before Hermes has done one useful job.

6:41Do not do that. Get one working path

6:44first. One model, one surface, one

6:46workflow, then expand. Now, let us talk

6:50about the part people mess up

6:51constantly, models. Hermes can use a lot

6:55of different providers. It can use

6:57OpenRouter, Anthropic, OpenAI, Codex,

7:01New Portal, Google, Deep Sea, Kimmy,

7:04Gwen, XAI, local endpoints, and so many

7:08more. Literally, the sky is the limit

7:10here, or rather, your wallet is the

7:13limit here. That flexibility is

7:15powerful, but it also creates a trap.

7:18The trap is thinking there's one perfect

7:20model for Hermes, because there is not.

7:23There are great cheap models if you're

7:26just exploring Hermes. There are great

7:28cheap models for background work that

7:30Hermes does while you're away, and there

7:33are great affordable or expensive models

7:36for coding sessions. But, here is the

7:39model ladder that I would use. For

7:42serious autonomous work, coding, tool

7:45use, and anything where you need

7:47reliability, I would start with the

7:49strongest model that you can afford.

7:52That usually means a top Claude, Sonnet,

7:56or Opus model, a strong OpenAI GPT or

7:59Codex model, or whatever the current top

8:03agentic model is in your provider list.

8:05At the time I'm filming this video,

8:08Hermes can route through providers like

8:10OpenAI, Codex, Anthropic, OpenRouter,

8:13New Portal, and the live model list

8:15changes constantly. So, I'm not going to

8:18pretend one model will be the winner

8:20forever. The rule is more important than

8:23the name. Use your strongest model when

8:25Hermes is changing files, writing code,

8:28using tools across multiple steps, or

8:31working on something where failure is

8:33expensive. Use cheaper models for

8:35summarizing, formatting, extraction,

8:38tagging, or routine background work. For

8:41example, if Hermes is summarizing a

8:43transcript, I do not need the most

8:44expensive model on Earth. I would rather

8:47route to something like Gemini, Flash,

8:51Deep Sea, Quen, Haiku, GPT Mini, or

8:54another cheap, fast model that is good

8:56enough for structured work. For long

8:58context research, I would look at models

9:01with very large context windows. Claude,

9:03GPT, Quen, Kimmi, Deep Sea, they all

9:06have options here depending on your

9:08provider. For privacy or learning, local

9:11models through Ollama or LM Studio are

9:14useful, but you need to be honest about

9:16the trade-off. Local models are great

9:19for simple private tasks,

9:20experimentations, and workflows where

9:23you do not want data leaving your

9:25machine. But, if you're asking Hermes to

9:27run a messy, multi-step workflow, a weak

9:31local model can waste more time than it

9:34saves. So, my practical setup is this.

9:38One strong default model for important

9:40work, one cheap fast model for

9:43background jobs, one long context model

9:46for giant documents and transcript work,

9:48one local model if privacy or cost

9:51matters to you, and if you're running

9:53Hermes heavily, use fallback providers

9:55or credential pools so one account

9:57failure does not kill the whole system.

10:00Do not ask what is the best model, ask

10:03what job is this model supposed to be

10:06good at doing. Two quick hacks I have

10:09here for you. Hack number one is there

10:12are a ton of subscription services like

10:14ChatGPT's Codex subscription where for

10:17$20 a month or $100 a month, you can

10:20plug your AI subscription into Hermes.

10:23So you don't have to pay for anything

10:25you weren't already paying for. And hack

10:28number two, you can literally tell

10:30Hermes, "Use this model when I'm doing

10:32heavy work like coding, tool calls, so

10:35on and so forth. Use this cheaper model

10:37when doing cron jobs and background

10:38work." You can tell Hermes when you'd

10:41like to use which model, and it will do

10:44the setting up for you. There's no need

10:45for you to be or get technical anywhere.

10:49Now, let's dive into memory. Now we get

10:52to the feature that makes Hermes feel

10:54different, memory. But I want to be very

10:57clear about something. Memory is not

11:00magic, and memory is definitely not let

11:03the AI remember every random thing

11:05forever and hope it becomes smarter.

11:08This is how you create a haunted

11:10assistant. Hermes has built-in memory

11:13that is intentionally small and curated.

11:16There are two core files Hermes uses for

11:19memory. user.md is for who you are,

11:23preferences, communication style,

11:25expectations, things like the assistant

11:28should know about you. And memory.md is

11:31for the agent's notes, environment

11:33facts, project conventions,

11:35lessons learned, workflows, things that

11:38help it operate better every day. These

11:41are injected into the system prompt at

11:43the start of every session you have with

11:45Hermes, which means they are fast and

11:48always available. But, they are limited

11:51on purpose. That is good if everything

11:54is memory, nothing is memory. The rule

11:57is simple. Memory should save facts that

12:00stop you from repeating yourself, not

12:02task progress, not temporary to-do list,

12:06not every random conversation. Good

12:08memory is something like the following.

12:12User wants compact Telegram replies with

12:15proof first. Bad memory is on Tuesday,

12:18we talked about maybe making a video

12:20someday and the user seemed interested.

12:23If it is a stable preference or

12:25environment fact, memory. If it is a

12:27procedure, skill. If it is a long

12:30research note, memory wiki. If it is a

12:32past conversation, then session search.

12:35That distinction is what stops memory

12:37from becoming slot. Now, if you want to

12:39go beyond built-in memory, Hermes now

12:42supports external memory providers. This

12:44is where things get interesting.

12:46Built-in memory stays active, but you

12:49can add on external provider for deeper

12:52recall, semantic search, knowledge

12:54graphs, and cross-session context. The

12:57command you could run on your terminal

13:00is bash Hermes memory setup, or you can

13:03literally ask it inside your chat to

13:06help you set that up. And you can check

13:08what is active, what isn't. Hermes

13:11supports providers like Honcho, Memo,

13:14Hindsight, Supermemory, and so many

13:17more. Only one external provider can be

13:20active at any point in time, and it is

13:22additive. It does not replace the

13:24built-in memory, which is great. Here is

13:26how I would think about memory

13:29providers. Memo is interesting if you

13:32want a dedicated memory layer for

13:34personalized AI agents. It focuses on

13:37extracting, storing, linking, and

13:39retrieving memories efficiently. Their

13:42newer memory system uses entity linking,

13:44keyword search, and temporal reasoning,

13:47which is exactly the kind of thing you

13:49want when an agent needs to remember

13:52people, projects, or and different

13:55changes over time. Honcho is interesting

13:58if you want user modeling and

14:00multi-agent [music] alignment.

14:02Basically, the agent builds a richer

14:05model of you [music] and the

14:06relationship across sessions. Hindsight

14:10is interesting if you care about

14:12knowledge graphs and entity

14:14relationships. That can be useful when

14:16you want the agent to connect people,

14:18projects, companies, and different

14:20decisions. And super memory and similar

14:23engines are useful if you want a larger

14:25memory and context layer that can scale

14:28beyond a few curated notes. The

14:31important part here is not the

14:33provider's name. The important part is

14:35the memory's architecture. I would use

14:38built-in memory for the tiny set of

14:41facts Hermes always needs. I would use

14:43an external provider like Memo or Honcho

14:46for richer personal and project recall.

14:49I would use a wiki and notes

14:52for source material, long research,

14:54scripts, and anything you want to

14:56inspect manually later on. And I would

14:59use the session search when I need to

15:01find something we discussed before, but

15:03I do not want it permanently injected

15:05into every prompt. That is the memory

15:09stack: tiny curated memory, external

15:12semantic memory, searchable session

15:15history, and human-readable wiki. This

15:18is how you make your Hermes actually

15:21compound. Now, let's get into something

15:24I personally love a lot, skills. The

15:27next piece of the puzzle is skills. This

15:30is where Hermes gets really powerful and

15:33also where people completely

15:35misunderstand it. Memory is for facts,

15:38skills are for procedures. If I correct

15:41Hermes one time and say do not write

15:44scripts like that, that might become

15:46memory. But if Hermes learns a

15:48repeatable process for writing scripts,

15:51researching competitors, creating a PDF,

15:54that becomes a skill. A skill is

15:56basically an SOP, a standard operating

15:59procedure that the agent can load when

16:02the task matches. It includes

16:05instructions, commands, pitfalls,

16:07templates, scripts, and examples. That

16:10means Hermes does not just remember that

16:13you like something, it remembers how to

16:15do the thing. For example, my YouTube

16:18agent does not just know Charbel likes

16:21to make YouTube videos. It has a whole

16:24entire process. It checks posted,

16:28rejected ideas, it checks VidIQ, it sees

16:31my rivals competitors transcripts, it

16:34studies my voice, writes, and figures

16:36out the SEO, and then it push his and

16:39verified. That is not a memory, that is

16:42a skill. That is the difference between

16:44an AI assistant that gets corrected

16:46forever and an AI assistant that

16:48actually improves. If you want real

16:51advantage with Hermes, do not just ask

16:54better prompts. Turn repeated work into

16:57skills. Now, let us talk about tools.

17:00Tools are how Hermes acts on the world.

17:03Without tools, Hermes is mostly

17:05conversation. With tools, it can read

17:08files, write files, run terminal

17:10commands, inspect images, generate

17:13audio, and so much more. That is

17:16powerful, but there's a line you need to

17:18respect. The goal is not to make Hermes

17:20reckless. The goal is to make Hermes

17:22useful. I want Hermes reading and

17:25preparing work all day. I do not want it

17:27sending emails, posting tweets for me,

17:30spending my money, or changing important

17:32systems without my approval. That is how

17:35I think about safety. Low-risk internal

17:38work can be fast. External or

17:40irreversible work needs a human

17:42checkpoint. For example, I'm totally

17:45fine with Hermes drafting five email

17:47replies. I'm not fine with it sending

17:50those replies without me reviewing them

17:52first, at least until the point that I

17:55have trained it enough and I'm

17:56comfortable enough with its output

17:59repeatedly. That is not anti-automation.

18:03That is how you keep automation useful.

18:05A good agent setup is not AI can do

18:08anything. A good agent setup is AI can

18:10do almost anything and all of the prep

18:14and it knows exactly

18:15where to stop. Let's move on to

18:17scheduled jobs.

18:19Now, this is the feature that makes

18:21Hermes stop feeling like a chat app,

18:24scheduled job. A chatbot waits for you

18:27to open it. A scheduled Hermes job shows

18:30up with work already done. That is a

18:32massive difference. Hermes has cron jobs

18:36and you can manage them with different

18:39create,

18:42Hermes cron run. You can even ask it

18:44just in the chat like, "Hermes, what are

18:46my crons? What is the list of my crons?

18:48Hey Hermes, can you run this cron

18:50manually right now? Can you pause this

18:52cron for me?" By the way, all the

18:54terminal commands I'm giving you

18:55throughout this video, you can also type

18:57inside a Hermes chat window using

18:59{slash} commands as well like {slash}

19:01cron for example, or just plainly asking

19:04it, "What are my crons?" In practice,

19:07this lets you build things like a

19:09morning brief, a daily YouTube

19:11opportunity scanner, a competitor

19:13monitor, a weekly content performance

19:15review. But, here's the thing people

19:18miss. Cron jobs run in fresh sessions.

19:21So, if you write a vague prompt like

19:23check everything and tell me what

19:26matters, that is a bad job. A good

19:28scheduled job is self-contained. It says

19:31what to check, where to check, what

19:33counts as important, what to ignore, and

19:36where to deliver that. For example,

19:39every morning at 8:00 a.m., check the

19:42last 24 hours of comments on my YouTube

19:45my YouTube's performance and competitor

19:47uploads and return only three things,

19:49one urgent issue, one content

19:51opportunity, and one recommended action.

19:54Keep it under 12 bullet points. That is

19:57much better because the rule is simple.

19:58Schedule things that help you make

20:01decisions faster and get more work done.

20:04Let's move on to sub-agents. Next up is

20:07sub-agents. Sub-agents are one of the

20:09features that sound like science fiction

20:11until you understand their actual use

20:14case. Uh sub-agent is useful when the

20:17work can be split cleanly. For example,

20:20if I'm researching a new video idea, I

20:23can have one sub-agent inspect

20:25competitor transcripts, one sub-agent

20:28check keyword demand, and one sub-agent

20:31review my existing content pipeline.

20:33Then, Hermes combines those results and

20:36makes a decision using all of the things

20:39the sub-agents did. That is extremely

20:41useful, but sub-agents are not magic

20:44employees. They do not automatically

20:46know everything. They need context. They

20:49need constraints. They need

20:51verification. The mistake is spawning

20:53five agents with vague instructions and

20:55then trusting the output like it came

20:57from a senior employee. That is not how

21:00this works. The right way is to give

21:03each sub-agent a narrow job. Something

21:05like read these three transcripts for me

21:08and extract the title promise hook

21:11structure and weak spots. That is where

21:14sub agents shine parallel work narrow

21:16scopes clear outputs and verified by

21:20main agent if you will. If you use them

21:23like that they are incredibly useful. If

21:26you use them like vague employees they

21:28become confusing. You can also ask

21:31Hermes to spawn sub agents for different

21:33tasks and if you use Hermes long enough

21:36you'll notice it will just spawn them

21:38for you whenever it decides that it

21:41needs to. Profiles are one of the most

21:44underrated Hermes feature. A profile is

21:47basically a separate Hermes home. Each

21:50profile can have its own configuration

21:52API keys memory cron jobs and

21:55personality. That means you can create

21:57specialist agents. You can create a

21:59coding profile a YouTube profile a

22:01business profile. This matters because

22:04one assistant should not know or do

22:07everything. My YouTube agent for example

22:09should know my content style and filming

22:12preferences. My coding agent should know

22:14my repos my coding standards and what

22:18are my deployment tools. Those should

22:20not all be the same brain. Profiles are

22:22how Hermes stops being one generic

22:25assistant and starts becoming a small

22:28team. You can run a command here like

22:30Hermes profile create research if you

22:32want to run it from the terminal or

22:35again ask it in the chat as simply as

22:38that. Then that profile gets its own

22:40alias. So if I create a profile called

22:43researcher I can open that researcher

22:45profile. I can enable or disable

22:48anything I want and customize it to my

22:50liking. You can by the way also do this

22:53on Hermes desktop over here. I have all

22:56of my different profiles and I can give

22:58each and every single agent its own

23:02different clean memory cleaner skills,

23:04cleaner permissions, and cleaner

23:07behaviors. This section matters because

23:10this is where a lot of Hermes videos get

23:13weak. They show you a setup, they show

23:15you a few features, and then they say

23:17something vague like, "Use it for your

23:19workflows." That is not very useful.

23:21This is why people install Hermes.

23:23People do not install Hermes because

23:25they want another chat app. They install

23:27it because they want one of seven

23:29things. The first reason is phone

23:32delegation. I'm away from my laptop, but

23:34I can still give it a real task to my

23:36agent that has my files, my tools, my

23:39memories, my skills, and my working

23:41environment. That is why so many Hermes

23:43videos are about Telegram, VPS setup,

23:46and 24/7 assistance. The second reason

23:50is long-running projects. With ChatGPT,

23:53you have a conversation. With a coding

23:55agent, you can push one task through a

23:57repo. With Hermes, the interesting thing

24:00is that you can give it a larger

24:02objective, break it into stages, use

24:05{slash} goal, assign sub-agents, and

24:08keep state as work moves forward. A

24:11useful skill for you to learn, by the

24:13way, here is the {slash} goal command.

24:15It will make Hermes keep working at a

24:17job until it meets the goal you

24:19outlined. A pro tip here is you can ask

24:22Hermes to help you create the best

24:24{slash} goal possible by telling it, "I

24:27want this done. Help me create the best

24:29{slash} goal possible." Literally that

24:31simple. The third reason people install

24:33Hermes is developer and ops work. This

24:36is one of the strongest Hermes use cases

24:38because Hermes can live next to your

24:40repo. It can use your terminal tools,

24:42and it can report back through your own

24:45phone. If you're a developer, this is

24:46probably one of the first serious

24:48workflows I would build. The fourth

24:50reason is exception monitoring. This is

24:53where scheduled jobs become genuinely

24:55useful. A lot of people hear cron job

24:58and immediately think of a a daily

25:00brief. Daily briefs are fine. They're

25:02They can be useful, but they're not

25:05usually the highest leverage version of

25:07a cron job. The better version is only

25:09wake me up when something changed that

25:11matters. For example, like if you have

25:13an online shop, you can schedule crons

25:16with Hermes that ping you when cost

25:18spikes, when app reviews land, when

25:21Stripe health declines, so on and so

25:23forth. The point is not that Hermes

25:25sends you more notification. The point

25:27is that Hermes can monitor a boring

25:30system for you. It can stay quiet when

25:33nothing matters and only spend model

25:35tokens when judgment is actually needed.

25:37The fifth reason people install Hermes

25:40is memory and knowledge systems. People

25:42are not just asking, "Does Hermes have

25:44memory?" They're asking whether that

25:47memory is actually useful. The hierarchy

25:49I would use for Hermes memory is

25:51built-in memory for high-signal facts

25:54and external memory providers when

25:56retrieval or cost agent memory becomes a

25:59bottleneck. The sixth reason is business

26:02operations. This is where you need to be

26:05careful because it can turn into fake AI

26:07employee nonsense really fast. But the

26:10real version of this are genuinely

26:12useful. Think of inbox triage, lead

26:15research, client follow-ups, customer

26:17review summaries, content gap analysis,

26:20weekly reports. The line is simple.

26:23Hermes should prepare work

26:25automatically, but humans should approve

26:28risky external actions first. So, the

26:30workflow is not let Hermes run your

26:32inbox. The workflow is every morning

26:35Hermes reads my inbox, separates signal

26:37from noise, identifies leads or urgent

26:40replies, drafts responses in my voice,

26:43attaches that to chat, and waits for my

26:46approval before sending. And finally,

26:48the seventh reason people install Hermes

26:51is content and market intelligence. Bad

26:54workflow here is if you're a content

26:56creator, find me AI news, for example,

26:59depending on your niche, of course. A

27:01good workflow is these are my

27:03competitors. Check them out every day.

27:05Pull their videos that are still gaining

27:07views. Read those transcripts. Compare

27:09them to my posted videos. Exclude what

27:12is already in my pipeline and recommend

27:14the one video that I should film next

27:17with your evidence as to why. That is

27:20exactly the kind of thing Hermes is good

27:22at because it can combine search,

27:24browser work, transcript, memory, and a

27:27final recommendation. Those are

27:29genuinely Hermes' strengths. So, if

27:32you're a creator, yes, Hermes can become

27:35a content intelligence system. That is

27:37the real pattern across all of these

27:40workflows. Do not install Hermes because

27:42it has features. Install Hermes when you

27:44have work that needs context, tools,

27:47memory, repetition, and a safe handoff

27:50back to you. And that leads us directly

27:54into the mistakes because the fastest

27:56way to ruin Hermes is to install it for

27:59the right reason and then set it up in

28:02the wrong order. Let's get started. Let

28:04me save you from the mistakes that make

28:07Hermes feel worse than it is. Mistake

28:09number one, adding too many tools way

28:12too early. If Hermes has access to

28:14everything before you know what you want

28:17it to do, you're just creating chaos.

28:19Start with the tools needed for your

28:22first workflow. Mistake number two,

28:24saving everything to memory. Memory

28:27should be curated. It should not become

28:29a dumping ground of everything and

28:31anything. Mistake three, using the

28:33cheapest model for hard work. If a task

28:36requires tool use, coding, or multiple

28:38steps, a weak model can waste more money

28:41in retries than a strong model would

28:43have cost you in the first place.

28:45Mistake number four, making every idea a

28:49cron job. A scheduled job should have a

28:51clear decision and output. If it does

28:54not, it just becomes notification spam.

28:56And this will just make your life a

28:59living hell every single day. Mistake

29:02five, trusting sub-agents without

29:04verification. Sub-agents are useful, but

29:06their summaries are still outputs that

29:08need checking. Mistake six, building

29:11profiles before workflows. Do not create

29:14five specialist agents before you know

29:16what those specialist agents are

29:18actually here to do. Mistake number

29:20seven, treating Hermes like a magic

29:22employee. Hermes is powerful, but it is

29:25still in need of systems, of good

29:27instructions, of good tools, good

29:29memory, good skills, good verification.

29:32That is the real game here. If you are

29:36starting today, here is the path I would

29:39follow. Day one, you're going to install

29:42Hermes, connect one model, and use

29:45Hermes desktop. Day two, you're going to

29:47connect Telegram and run one useful task

29:50from your phone. Day three, create one

29:53skill from a repeated task. Day four,

29:56create or connect, rather, one tool that

29:59actually matters for your work. Day

30:01five, create one scheduled jobs that

30:04returns a decision, not some data dump.

30:07Day six, try one one's agent workflow

30:10for research and review. And day seven,

30:13create your first specialist profile.

30:16That is enough. You do not need the

30:18perfect setup. One assistant that can do

30:21one real job, then another, then

30:24another. That is how Hermes compounds.

30:27The reason Hermes is exciting is not

30:29that it has a desktop app or Telegram or

30:32memory. Those are just features. The

30:35reason Hermes is exciting is that all of

30:37those parts that can connect into a

30:40single system. A system that remembers

30:43what matters. A A system that learns

30:45from your workflows, that shows up

30:47before you even ask for it. That is the

30:50difference between using AI as a search

30:52box and using AI as an operating layer.

30:56If you are new, do not try to master

30:58every feature today. Install it, connect

31:01one model, build one workflow, then turn

31:03the parts that repeat into skills,

31:06memories, schedule jobs, and profiles.

31:09That is how you learn 95% of Hermes

31:12agent. And if you enjoyed this video,

31:14make sure to subscribe because I have a

31:16ton more content coming your way. And

31:19comment Hermes if you'd like more videos

31:22about Hermes. Anyway, I hope you have a

31:24great day and I'll see you soon.