YouTube2Text

7 Hermes Agent Upgrades You NEED To Implement Today — Transcript

by Komputer Mechanic · 2,056 words · 288 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Hermes Agents is one of the most
  2. 0:01powerful AI tools you can use [music]
  3. 0:03right now. But for most people, it just
  4. 0:05feels like a chatbot inside Telegram.
  5. 0:07I've been running Hermes every day for
  6. 0:09months now. Not one agent, a whole crew
  7. 0:11of them. And I've learned what actually
  8. 0:13makes a difference. So in this video,
  9. 0:15I'm going to show you seven tips that
  10. 0:16takes Hermes from a bot you chat with to
  11. 0:18a team that works for you. And stay to
  12. 0:20the end because tip number seven is my
  13. 0:22favorite one. Just so you know, none of
  14. 0:24the tools that I mention in this video
  15. 0:26are sponsors. These are just things I
  16. 0:28genuinely use. All right. So the first
  17. 0:30one, build a crew, not one agent. Most
  18. 0:33people use a single agent for
  19. 0:34everything. Research, writing, coding,
  20. 0:37all in one chat. And after a while, it
  21. 0:39starts mixing things up. So I split the
  22. 0:41work. For example, I have a main agent
  23. 0:43called Orchestrator who plans and runs
  24. 0:45the show. Then research finds things
  25. 0:47out, writer writes, and developer
  26. 0:49builds. Each one [music] has a defined
  27. 0:51job. And you don't need a terminal to
  28. 0:53set these things up. Just ask your main
  29. 0:54agent something like, "Create a new
  30. 0:56persistent agent called research. It's
  31. 0:58only job is to find things out and
  32. 1:00summarize what it finds. Give it its own
  33. 1:02instructions and its own memory. And
  34. 1:04tell me when it's ready." And that's it.
  35. 1:06Your main agent builds it for you.
  36. 1:09Repeat it for every agent you want. Or
  37. 1:11ask your main agent to create several of
  38. 1:13them at once. Now here's my favorite
  39. 1:15part. Make one Telegram group and turn
  40. 1:17on topics in the group settings. Now you
  41. 1:19can give every agent its own topic, like
  42. 1:21its own channel. So you talk to research
  43. 1:23directly, to writer directly, and the
  44. 1:25same for every agent. And you always
  45. 1:27know who is doing what. And because
  46. 1:29every topic is its own session, every
  47. 1:32agent keeps its own conversation. And
  48. 1:34you don't end up blowing up one context
  49. 1:36window like you would do when talking to
  50. 1:37all of them in one chat. The tricky
  51. 1:39part, out of the box, Hermes sends every
  52. 1:42message in that group to your main
  53. 1:44agents. But there's a simple workaround,
  54. 1:46a small plugin that sends each topic to
  55. 1:48the right agent. And I have a prompt for
  56. 1:50you for that in the description section.
  57. 1:52So now you've got a crew, but every
  58. 1:54agent in that crew runs on an AI model.
  59. 1:56[music] And models cost you money.
  60. 1:58Bringing us to tip number two, you don't
  61. 2:00need the most expensive model. Most
  62. 2:02people pick the most expensive model for
  63. 2:04everything, but I recently had a whole
  64. 2:06memory system for my AI agents built by
  65. 2:08MiniMax M3, which cost a lot less than
  66. 2:11the top models from Anthropic and
  67. 2:13OpenAI. Cheaper doesn't mean bad. It
  68. 2:15builds the whole thing. Every agent can
  69. 2:17run its own model. So, think about what
  70. 2:20each one does. A developer agent writing
  71. 2:22code, sure, give it a stronger model.
  72. 2:24But an agent that is probably just
  73. 2:26checking your email or summarizing an
  74. 2:27article probably doesn't need the most
  75. 2:29expensive model in the world. Put it on
  76. 2:31a cheaper one. And even inside one
  77. 2:33provider, there's a whole family of
  78. 2:35models. OpenAI, for instance, has big
  79. 2:37models like GPT-Xtra and GPT-5.6 and the
  80. 2:40cheaper ones like GPT-5.5 or GPT-5.4
  81. 2:43mini. Same company, same account. You
  82. 2:46don't always have to pick the top one.
  83. 2:48And even with the same model, there are
  84. 2:50different effort levels that you can
  85. 2:51set. You can go from light all the way
  86. 2:53to extra high. So, think about the
  87. 2:55effort that the agent needs for that
  88. 2:56particular task and adjust the model and
  89. 2:59adjust the effort level as you see fit.
  90. 3:01Now, here's something a lot of people
  91. 3:02probably don't know. If you already pay
  92. 3:04for a ChatGPT plan, you can sign in to
  93. 3:06Hermes with that account and use the
  94. 3:08OpenAI models without paying anything
  95. 3:10extra. And the same applies to Grok. If
  96. 3:12you have a super Grok or X subscription,
  97. 3:15the higher your plan of course, the more
  98. 3:16usage you get. And that usage typically
  99. 3:18works out a lot cheaper than paying as
  100. 3:21you go through the API. So, now your
  101. 3:23crew is cheaper to run. But there's one
  102. 3:25problem every crew runs into. [music]
  103. 3:27They forget things. So, tip number
  104. 3:29three, give your agents a real memory.
  105. 3:31Agents don't have long-term memory. Chat
  106. 3:33history fills up and when it does, the
  107. 3:35agent has no option than to summarize it
  108. 3:37to make room. This is what is called
  109. 3:39compaction and every time it compacts,
  110. 3:41details get lost. It's like a whiteboard
  111. 3:44that someone keeps wiping. Say tell your
  112. 3:46agents on Monday exactly how you like
  113. 3:48your reports written. Probably by Friday
  114. 3:51after a few long chats, that's probably
  115. 3:52gone. It gets summarized away. So, I
  116. 3:55give my crew a wiki, plain notes in
  117. 3:57plain markdown files on my own server.
  118. 4:00Every agent resets before it answers and
  119. 4:02writes to it after it works. And because
  120. 4:05they are plain markdown files, I can
  121. 4:06open any page, read it to learn what the
  122. 4:08agents are documenting, and also fix
  123. 4:10where I find inconsistencies myself.
  124. 4:12[music] But hey, you can also outsource
  125. 4:15this to a librarian agent like I did in
  126. 4:16my setup, where the agents will take
  127. 4:18care of this for you on an
  128. 4:21The best part is that it's not only your
  129. 4:23Hermes agent, any AI agent you use like
  130. 4:25Claude Code or Codex can read the same
  131. 4:27wiki and be up to speed in second. And
  132. 4:30one rule, keep your own facts separate
  133. 4:32from what the agents read online. Your
  134. 4:33decisions and your clients live in one
  135. 4:36place and articles they read live
  136. 4:38somewhere. Because someone else's
  137. 4:40marketing claim should never become your
  138. 4:42fact. This is inspired by Andrej
  139. 4:44Karpathy's LLM wiki concept. So, now
  140. 4:46your crew remembers everything, but so
  141. 4:48far it only works when you talk [music]
  142. 4:50to it. Bringing us to tip number four,
  143. 4:52let your agents work without you. Hermes
  144. 4:55can run real tasks on a schedule.
  145. 4:56[music]
  146. 4:57That's called a cron job. Not reminders,
  147. 4:59real work. For example, every morning at
  148. 5:018:00 a.m. my agent sends me what is
  149. 5:03called a morning brief and that contains
  150. 5:05a summary of my unread emails, the tasks
  151. 5:07I need to handle, my calendar, and the
  152. 5:09weather for that day. Basically, it
  153. 5:11preps me for the day ahead. I don't ask
  154. 5:13for it. It's just waiting on my phone
  155. 5:15when I wake [music] up. Or say you have
  156. 5:17an agent check your email every 30
  157. 5:19minutes and maybe draft a reply for you.
  158. 5:21Now here's a small trick. Most of these
  159. 5:23checks probably find nothing new. So,
  160. 5:25without this trick, your phone buzzes
  161. 5:27every half an hour just to tell you
  162. 5:29nothing happened. So, in every scheduled
  163. 5:31job, tell the agents if there is nothing
  164. 5:33new to report, just stay quiet. Hermes
  165. 5:35has a built-in way to do that, so
  166. 5:37nothing gets sent at all. I will also
  167. 5:39put the exact prompt for this in the
  168. 5:41description section. But for your agents
  169. 5:43to be able to check your emails, your
  170. 5:45agents actually needs a tool to give it
  171. 5:47access to that. Which brings us to tip
  172. 5:49number five, connect your tools the safe
  173. 5:51way. Now, to connect my agent to Gmail,
  174. 5:53my calendar, Notion, and other apps that
  175. 5:56are out there, I use Composio. It has a
  176. 5:58huge pool of tools you can connect to.
  177. 6:00And there's a free plan to start with.
  178. 6:02And surprisingly, you can use it for so
  179. 6:04much without even having to pay. You
  180. 6:06sign in once in your browser and connect
  181. 6:08the tools you want your agent to use.
  182. 6:10And once you connect your Hermes agent
  183. 6:11to it, the agent gets a key to get
  184. 6:14access to these tools. And you can
  185. 6:15cancel it anytime. The best part is that
  186. 6:17your password access to these [music]
  187. 6:19tools never touches your server. Think
  188. 6:21of it like giving a house sitter a spare
  189. 6:23key, which you can take back whenever
  190. 6:24you want. And they never had the
  191. 6:26original key. Now, three rules I follow.
  192. 6:28First, give access to one agent only,
  193. 6:31not the whole crew. A developer agent
  194. 6:33writes codes. It probably has no reason
  195. 6:35to read your inbox. So, it shouldn't be
  196. 6:37able to. Second, start read [music]
  197. 6:39only. My finance dashboard, a new
  198. 6:41tutorial which is coming soon, reads my
  199. 6:42receipts straight from my Gmail. And it
  200. 6:44can never send an email or delete one.
  201. 6:46[music] It only needs to read, so
  202. 6:48reading is all it can do. You can always
  203. 6:51give more access later, but you can't
  204. 6:52take back an email that has already been
  205. 6:54sent or deleted. And the third tip, tell
  206. 6:56your agents that email is information,
  207. 6:59never instructions. Anyone can send you
  208. 7:02an email, and anyone can hide
  209. 7:03instructions inside one hoping your
  210. 7:05agents will follow them. This is what is
  211. 7:08called prompt injection. So, your agents
  212. 7:10reads what that email is, but never
  213. 7:13takes orders from it so that no one
  214. 7:15pulls some information from your agents
  215. 7:17through any tricky means. So, now you've
  216. 7:19got one crew connected to your tools.
  217. 7:21But what if you need more than one?
  218. 7:23Bringing us to tip number six, run
  219. 7:25several Hermes setups on one server. Say
  220. 7:28you want separate crews, one for content
  221. 7:30creation, one for work, one for
  222. 7:32personal, or one for each client you set
  223. 7:34agents up for. You don't need a separate
  224. 7:36server or system for each of them. With
  225. 7:39Docker, one server can run several
  226. 7:41completely separate Hermes setups,
  227. 7:43[music]
  228. 7:43and each one has its own isolated Hermes
  229. 7:46workspace, meaning its own agents, its
  230. 7:48own memory, and its own Telegram bot. No
  231. 7:51crew ever sees another crew's notes, and
  232. 7:54you only pay for one server. Now, at
  233. 7:56this [music] point, there's a lot going
  234. 7:58on, and I constantly keep jumping
  235. 8:00between Telegram folders and the
  236. 8:02terminal. [music]
  237. 8:02And that brings me to tip number seven.
  238. 8:04This is a big one. Build your own
  239. 8:06mission control. [music]
  240. 8:07I built one screen for all of it. I call
  241. 8:10it Nexura. I launch a mission and
  242. 8:13execute a plan set. Nothing runs until I
  243. 8:15approve it. And that's exactly why I'm
  244. 8:17comfortable letting it run. Then I watch
  245. 8:20the agents pass the work between
  246. 8:21themselves, and the results lands
  247. 8:23directly in docs, where I can find it
  248. 8:25next week instead of scrolling back
  249. 8:27through chats. I can chat with any agent
  250. 8:29from here. There's an agent city that
  251. 8:30lights up when an agent is working. So,
  252. 8:33one glance tells me who is busy. [music]
  253. 8:35And hey, I also set up an agent's login
  254. 8:37database. Whenever an agent performs a
  255. 8:39task, it logs in the database a summary
  256. 8:42of the task that they performed, the
  257. 8:43time that they performed that task, the
  258. 8:45AI model that they used, and [music] the
  259. 8:46agent's name. Now, this way, I can
  260. 8:48monitor everything that the agents
  261. 8:51[music] are doing. Additionally, I can
  262. 8:53check my machine's health and see every
  263. 8:55scheduled job, including the 8:00 a.m.
  264. 8:57brief [music] that I talked about. And
  265. 8:59with the voice mode, I can just talk to
  266. 9:01the agents, and they can also talk to me
  267. 9:03back if I don't feel like typing. The
  268. 9:04whole mission control dashboard is
  269. 9:06private. It runs on my own machine, and
  270. 9:08only my devices that I own can reach it.
  271. 9:11Because think about what your crew knows
  272. 9:13about [music] you, your emails, your
  273. 9:14clients, and that is why security should
  274. 9:16be your highest feature. [music] Now,
  275. 9:18before I was chatting with a bot, but
  276. 9:20now I'm running a crew of AI agents. You
  277. 9:23can build it yourself from all the
  278. 9:24prompts that I'm going to provide in the
  279. 9:25description section, or you can install
  280. 9:27it with an auto installer that I built
  281. 9:29for you guys, saving you all the AI
  282. 9:31tokens and time it would cost you to
  283. 9:32build this. Now, if you found this video
  284. 9:34helpful, hey, hit the thumbs up button
  285. 9:36and subscribe to my channel to see more
  286. 9:38like this. And with that, thanks for
  287. 9:39watching this video and I'll see you in
  288. 9:41the next one. Bye.

About this transcript

This page contains the full transcript of 7 Hermes Agent Upgrades You NEED To Implement Today by Komputer Mechanic, generated from the public captions YouTube serves with the video. The transcript has 2,056 words across 288 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.