How I Use Hermes Agent for 95% Cheaper (Feels Illegal) — Transcript

Full transcript

1. 0:00If you're using Hermes agent, you're
2. 0:01probably burning through a ton more
3. 0:03tokens than you probably realize. And
4. 0:06I'm talking hundreds of dollars in API
5. 0:09credits. And the crazy part is a lot of
6. 0:10that money isn't paying Hermes to do
7. 0:12more work or give you a better answer.
8. 0:14It's getting burned on context,
9. 0:16instructions, and tools that Hermes
10. 0:18keeps sending to the model every time
11. 0:20you use it. So, in this video, I'm going
12. 0:22to show you the setup that I use to cut
13. 0:24Hermes cost by up to 95%
14. 0:27on the right jobs without sacrificing
15. 0:30the quality where it matters. I'll show
16. 0:32you how I manage the context, choose the
17. 0:34right model for each job, and set a few
18. 0:36simple rules that stop Hermes from
19. 0:38wasting tokens. So, let's start with
20. 0:40what Hermes is actually sending every
21. 0:43time you message it because that's where
22. 0:45most of the waste begins. So, to save
23. 0:47money on tokens, the key is finding the
24. 0:50waste because every request has an input
25. 0:54side and an output side, and you pay for
26. 0:56what the model reads and what it writes.
27. 0:59And the part that people miss really is
28. 1:01the input. So, your your message can
29. 1:03arrive with the system prompt,
30. 1:05conversation history, memory, files,
31. 1:08tool instructions, and a whole bunch
32. 1:10more. And so, this is how a simple
33. 1:13one-line prompt turns into tens of
34. 1:16thousands of tokens being used before
35. 1:19Hermes has done a single thing. So, you
36. 1:22could see here, this is the token
37. 1:23optimizer dashboard, and you can see
38. 1:25here my my overhead is 40,000 tokens.
39. 1:29All right. So, we're going to walk
40. 1:31through this in in just a moment, and
41. 1:33I'm going to walk you through some steps
42. 1:34of how you can bring this down. And
43. 1:36also, how to optimize how you interact
44. 1:39with these agents because there are some
45. 1:41specific things that you need to do in
46. 1:43every single session that will reduce
47. 1:45your token usage massively and also
48. 1:48increase the accuracy and reliability of
49. 1:51the model that you're using. So, we
50. 1:52don't need to delete every single tool
51. 1:55because installed doesn't actually mean
52. 1:57that it's fully loaded every single
53. 1:59time. So, you need to check what Hermes
54. 2:02is actually sending when you put in that
55. 2:04first message. It's also a pretty smart
56. 2:07decision to check your skills and MCPs.
57. 2:11It's worth just cleaning those up to
58. 2:13remove any bulk and unnecessary token
59. 2:15usage. Checking your system prompt as
60. 2:17well, we're going to go through that in
61. 2:18just a second as well. And also cleaning
62. 2:20up the context, this is a huge hack
63. 2:23right here. And this is what I learned
64. 2:24through the Claude Code Certified
65. 2:26Architect program that I passed the
66. 2:28examination for a couple of weeks ago is
67. 2:31clearing your context, right? Don't wait
68. 2:33until your context window is 60, 70, 80,
69. 2:3690% full. You want to be clearing your
70. 2:39context window at about 40, 50% maximum.
71. 2:42And the reason why we do that is because
72. 2:44the model will start to lose information
73. 2:46in the middle. Because remember, every
74. 2:48request that you send to the model
75. 2:50includes the entire conversation that
76. 2:54you have in your current chat. So, to
77. 2:56improve the accuracy and just general
78. 2:58intelligence of the model, what I
79. 3:00usually do is create a handoff MD file
80. 3:03when I'm about 40% in, clear the
81. 3:05session, and start from there. And that
82. 3:07that's different to compaction.
83. 3:09Compaction is going to compress your
84. 3:12session. You don't have any say in what
85. 3:14gets saved, and that can be a problem.
86. 3:16But, if Hermes is supposed to work like
87. 3:18an actual agent, it also needs to be
88. 3:20available 24/7 when your laptop's
89. 3:22closed. And that is where Hosting a Fix
90. 3:25into this setup. You see, I run Hermes
91. 3:27on Hosting a VPS, and that's what keeps
92. 3:30it online when my laptop is closed. You
93. 3:32can see here, this is a managed
94. 3:34installation. And if we scroll down, you
95. 3:36can see some of the benefits you get
96. 3:37from having it fully managed. You don't
97. 3:39have to worry about any of the security,
98. 3:43Telegram pairing is already built in,
99. 3:45and it's literally just one-click setup
100. 3:47and super easy and quick to use. And
101. 3:49this also includes like weekly backups,
102. 3:51manual snapshots, which is useful if you
103. 3:54want to update Hermes or change the
104. 3:55setup. And the way that you get started
105. 3:57is really simple. You click get started,
106. 4:00you pick the plan. We're going to grab
107. 4:01this one right here. We're going to move
108. 4:03through to the cart page. And you got
109. 4:05some options here, right? You can sign
110. 4:07up for 24 months, 1 month, 12 months.
111. 4:09This is going to give you the biggest
112. 4:10discount. You're selecting Hermes agent
113. 4:12here, and then you're going to pop in my
114. 4:14discount code where it's going to get
115. 4:15you an additional 10% off. And the code
116. 4:18is growth lab 10. We're going to apply
117. 4:22that discount. And boom, we're saving a
118. 4:24bit of cash there. And then you'll just
119. 4:25continue with your purchase. And you can
120. 4:27decide if you want to use extra add-ons
121. 4:29or not. But if you want to use Hermes
122. 4:3124/7 in Telegram like I've got here,
123. 4:34you're going to need it hosted on a VPS
124. 4:36somewhere. And hosting it is just one of
125. 4:38those options. I use it, I have no
126. 4:39issues with it, and it works extremely
127. 4:41seamlessly with Telegram on my phone and
128. 4:43also in my terminal on my laptop.
129. 4:45Anyway, thanks for hosting her, for
130. 4:46sponsoring this video. Okay, so let's
131. 4:48talk about token optimizer. This is a
132. 4:50free GitHub plugin that you can install.
133. 4:54I've used this both on Hermes and Cloud
134. 4:56Code as well.
135. 4:58And if we scroll down, you can see some
136. 5:00of the features. It was ranked as a top
137. 5:0210.
138. 5:03You can use it on all of these different
139. 5:05agents. And here's what you get. Now,
140. 5:08I'm not going to run you through all of
141. 5:09this. What you can just simply do is
142. 5:11just grab the link, which is going to be
143. 5:13in the description, pop it into Hermes,
144. 5:15say, "Hey, can you run this?" It'll run
145. 5:17it and it'll give you a report looking
146. 5:19something like this.
147. 5:21But we could see here we've got smart
148. 5:22smart compaction, session continuity,
149. 5:25active compression, quality score,
150. 5:27session database, progressive
151. 5:28disclosure, and all this stuff. Model
152. 5:31routing nudges is kind of interesting.
153. 5:33And you can run the full uh audit using
154. 5:35{forward slash} token optimizer. You can
155. 5:37look at your 30-day trends. So, I'll
156. 5:39I'll leave it up to you. Um
157. 5:41the cool thing is that you can get the
158. 5:43status bar. So, over in Cloud Code in
159. 5:46the terminal, you can see the bottom
160. 5:49here, I've got this status bars, I can
161. 5:51see where my context is. Now, with uh
162. 5:54Hermes or Hermez, I don't even know how
163. 5:56to say it, you can see that we've got
164. 5:57the bar at the bottom here. I am in the
165. 5:59terminal setup. Um you can also just do
166. 6:01{slash} context and it'll give you this
167. 6:04command here. Same on the desktop, you
168. 6:06can just write context there. And one
169. 6:08way to check how many tokens you're
170. 6:10using on that first prompt, which is
171. 6:13where it's going to load your skills,
172. 6:14the the front matter, it's going to load
173. 6:16system prompts, and a bunch of stuff.
174. 6:18So, you can just pop in a simple prompt
175. 6:21like that, and you can see here, it's
176. 6:23what I've done here, and you can see
177. 6:25I've used 7%,
178. 6:27right, which is 18.9k of a 272k context
179. 6:31window. So, I I could run context as
180. 6:33well, and we can see what's being used.
181. 6:36So, this is kind of a quick and dirty
182. 6:38way of doing it if you don't want to do
183. 6:39a full audit using that tool. You can
184. 6:41see system prompt is taking up 2.6%
185. 6:45and the tool definitions is taking up
186. 6:473.7%. So, if we can delete some tools
187. 6:50that we don't even use, that's going to
188. 6:52free up some tokens. And then if we can
189. 6:55edit our system prompt, we can also save
190. 6:58some tokens as well. And you can see
191. 7:01down here at the bottom, and you can see
192. 7:02the token usage of these additional
193. 7:04context files here, but most of these
194. 7:06are pretty small anyway. But if we do
195. 7:08the same thing in uh in the desktop app,
196. 7:11you can see the same thing here. Now,
197. 7:14this model I'm using is MiniMax M3. It's
198. 7:17got a million tokens context window. So,
199. 7:19obviously, it's going to be a smaller
200. 7:21percentage, right? So, yeah, you can
201. 7:23install this. I think it's pretty pretty
202. 7:25worthwhile. I first did this with Claude
203. 7:26Code, and I was able to save a bunch of
204. 7:28tokens. So, you can see here in the
205. 7:30earlier on, it said 40k, now it's 19.2.
206. 7:33And if we click through to savings, we
207. 7:35can see how much money we might save if
208. 7:37we're paying for our usage. We've got
209. 7:39quick wins. Now, there's not really a
210. 7:41lot here. So, we can turn these on if we
211. 7:44want. It's also interesting to see which
212. 7:46tools are taking up the most tokens. And
213. 7:49we can see here delegation and
214. 7:51connections. I don't even know what they
215. 7:53are. So, definitely worth looking into.
216. 7:55Here we can review the Hermes skill
217. 7:57index. So, if we click on that, so you
218. 7:59can see here we're probably not going to
219. 8:00save too many tokens on that. And we've
220. 8:03got the persistent memory and profile
221. 8:05text. So, again, some other ways that we
222. 8:08can save some tokens. We've got habits
223. 8:10here, which I mentioned earlier, trends,
224. 8:13health, coach, token health score. And
225. 8:16you can see here issues detected many
226. 8:18unused skills. And this might be the
227. 8:20case for you because Hermes comes
228. 8:22pre-installed with a bunch of skills.
229. 8:25When you go through the setup, you know,
230. 8:27you can tick the ones to install. And
231. 8:29you can often end up installing skills
232. 8:31that you will actually never use. If we
233. 8:33come down to this manage tab at the
234. 8:34bottom, we can see some of the other
235. 8:37options that were mentioned in the
236. 8:38overview with uh active compression. You
237. 8:41can read through what what this does.
238. 8:43This quality nudge is just going to warn
239. 8:44you when your context quality drops
240. 8:47suddenly cuz it's actively monitoring.
241. 8:50So, if we go over to Claude code, for
242. 8:52example, we can see here the context
243. 8:53quality got an A for that. So, that's
244. 8:56that's good to go. And you can see we're
245. 8:57at 25%. So, no worries on that session.
246. 9:00Now, I don't see that inside of the
247. 9:02terminal here for Hermes, unfortunately.
248. 9:04There's a There are other settings like
249. 9:07loop detection, so detecting when the AI
250. 9:09is stuck retrying the same thing. That
251. 9:11That is going to burn through your
252. 9:12tokens. So, we've got delta mode. So,
253. 9:14what this does when the AI rereads a
254. 9:17file after editing it, it's actually
255. 9:18better not to read the whole thing. It's
256. 9:20token intensive. Instead, you can just
257. 9:22replace what went wrong with the new
258. 9:25information, copy whatever it is that
259. 9:27you're trying to replace. We've got a
260. 9:28structure map measurement, logs
261. 9:30compression events. Again, this is stuff
262. 9:32that you probably don't need to worry
263. 9:33too much about, but but worth having a
264. 9:35read to see if it can apply to you. It's
265. 9:37given us 15% here, but but as you can
266. 9:40see, all of these, six out of seven are
267. 9:42turned on. So, I'm just going to flip
268. 9:44So, if you want to work with the
269. 9:45structure map management, you can see
270. 9:46you need to put this in the terminal.
271. 9:47So, great tool, definitely check it out.
272. 9:50And let's move on to the next thing that
273. 9:51I want to show you, which is Ponytail.
274. 9:53So, you may or may not have heard about
275. 9:54this. This makes your AI agent think
276. 9:56like the laziest senior developer in the
277. 9:59room. So, this is essentially giving you
278. 10:01the 80/20 rule on your code, right?
279. 10:04Agents have the tendency to write a lot
280. 10:07of code for something that doesn't need
281. 10:10all of the lines of code that it's
282. 10:12written for you. So, I'm thinking here
283. 10:1480/20, right? You only need 20% that's
284. 10:16going to do 80% of the heavy lifting,
285. 10:18right? And you can see this was a number
286. 10:20one repository at one point. I've got
287. 10:22this installed for Claude Code and
288. 10:24Hermes. And you can see here the same
289. 10:25agent, the same tasks. Ponytail helps
290. 10:27your agent do it with less lines of
291. 10:29code, with less output tokens, a lower
292. 10:32cost, and it's faster. So, I don't know
293. 10:35anybody who wouldn't want those
294. 10:37benefits. So, again, you can just grab
295. 10:40the link. It's going to be in the
296. 10:41description. You can throw that into
297. 10:43your Hermes or whatever agent you're
298. 10:44using, and boom, you'll get it
299. 10:46installed, and it'll just be working in
300. 10:48the background without you needing to do
301. 10:50a thing. Now, it does come with a bunch
302. 10:52of commands here, so we can set the
303. 10:54intensity, or we can turn it off. So,
304. 10:56that basically means how heavily
305. 10:58Ponytail is going to be involved in the
306. 11:00code writing to reduce the amount of
307. 11:02code that's written. We can have
308. 11:03Ponytail review our code. We can audit
309. 11:06the repo, and then there's a few others
310. 11:08here that I haven't personally used. So,
311. 11:10some other things we want to take a look
312. 11:12at is the agents.md file. We want to see
313. 11:13if we can optimize that. Then there's a
314. 11:15few little cool things that I want to
315. 11:17add that you may or may not have heard
316. 11:18about. Hey, can we review my agents.md
317. 11:21file? Okay, so I want to review my
318. 11:23agents.md file. That's another area
319. 11:25where we can make some savings. If
320. 11:27you're using Claude, then you can also
321. 11:29review your claude.md file. That can
322. 11:32often get bloated where you add more and
323. 11:34more instructions and there's a point
324. 11:36where you need to do an audit and see if
325. 11:39you can redirect some of that
326. 11:41information. So, I do is I have some
327. 11:43path names for different files like my
328. 11:45voice.md. I have different routes inside
329. 11:48my Claude.md file. That file shows
330. 11:50Claude where to look. We don't include
331. 11:52that in the Claude.md or the agents.md
332. 11:55file. So, we can see here after asking
333. 11:57Hermes to review our agents.md file. We
334. 12:00see it's pretty good. Like it's not
335. 12:01particularly long. I've loaded it up
336. 12:03here in the right-hand panel. And there
337. 12:05is a few unnecessary things in here like
338. 12:07mentioning the Claude.md, mentioning the
339. 12:09codex.md. So, we could tighten that up.
340. 12:12But that's that's that's literally all
341. 12:13I've got here. So, in my case, honestly
342. 12:15not really a lot to change or remove
343. 12:18here. There's definitely space for
344. 12:20adding more because if you look at my
345. 12:21Claude.md file, you can see it's
346. 12:23actually quite long. And it's long for a
347. 12:26reason because it's the primary file
348. 12:28that loads for every single Claude code
349. 12:30session that I do inside my AI operator
350. 12:32OS. You can see you've got this quick
351. 12:34directory reference, which is what I was
352. 12:36mentioning earlier about routing to
353. 12:38different folders in my operator OS. You
354. 12:41can see we've we've got the the rules
355. 12:43for my voice.md, workflow discipline,
356. 12:46efficiency rules, workspace structure,
357. 12:48product and funnel rules, things like
358. 12:50this. So, maybe maybe I could optimize
359. 12:52that and condense those sections down
360. 12:54and have separate MD files for different
361. 12:56sections and have Claude apply those
362. 12:59whenever it needs to. So, I'm going to
363. 13:01add some more rules here and I'm going
364. 13:03to pop these in my notepad here. So,
365. 13:06prefer CLIs over MCP tools when both do
366. 13:10the job. Right? So, we're thinking here
367. 13:12GitHub MCP or GitHub CLI. Uh Google
368. 13:15Workspace MCP or Google Workspace CLI.
369. 13:19CLIs are designed for computers to use.
370. 13:23And in my opinion, they work faster,
371. 13:24they use less tokens. The second rule
372. 13:26here is move heavy exploration. So, when
373. 13:28it's actually reading many files to sub
374. 13:31agents, we don't want to use up our main
375. 13:33chat context for file exploration. So,
376. 13:36spin up sub agents for that. Definitely
377. 13:38going to help you save tokens in your
378. 13:39context window, and then we keep only
379. 13:41the summary in the main window. So, your
380. 13:43sub agent's going to do the exploration,
381. 13:45send the summary to the main chat, and
382. 13:47then the main chat is going to use that
383. 13:48summary, which is going to contain far
384. 13:50less tokens. So, the next rule we're
385. 13:52going to add to our MD file is
386. 13:54communication. This is going to make
387. 13:56sure our responses are as short and
388. 13:58concise as possible, which is going to
389. 14:01reduce our output tokens. So, always
390. 14:03talk in ASD STE100.
391. 14:06This is simplified technical English,
392. 14:09and say only what needs to be said.
393. 14:11Report only the elements needed for me
394. 14:13to make the right decisions, explained
395. 14:16clearly. This one's huge, because if
396. 14:18you're used to getting pages of response
397. 14:22from just small prompts inside of your
398. 14:25agent, this is going to fix that. And
399. 14:27it's going to have the added benefit of
400. 14:29saving you a whole bunch of tokens. So,
401. 14:31the other thing, and final thing, which
402. 14:33I think is probably the most important
403. 14:35thing to remember when you're working
404. 14:37with an agent, is your model selection
405. 14:38and your effort level. So, in Hermes, we
406. 14:41can use a bunch of the different open
407. 14:43source models. Now, I'm not here to tell
408. 14:45you which is better than the other. What
409. 14:47I do know is that Kimiko 3, Minimax M3,
410. 14:50Deep Seek V4 Pro pretty damn good
411. 14:53models. And you can hook up Open Router
412. 14:55with your Hermes instance, so you have
413. 14:57access to all of these models. And you
414. 14:59will have to pay for your usage for
415. 15:01these. You can also connect it to your
416. 15:03Cloud Code or Codex account, and use the
417. 15:06tokens that you get with your plan. So,
418. 15:08if we scroll up here, we should be able
419. 15:09to find ChatGPT or Codex subscription,
420. 15:12and we can use any of these models here,
421. 15:14which are just going to use up tokens
422. 15:15from our subscription. And the same here
423. 15:17with Anthropic, it's just going to use
424. 15:19up tokens in our subscription. So, let's
425. 15:22just say you want to use the models that
426. 15:24you have available with your
427. 15:25subscriptions, that's totally fine.
428. 15:27Okay, what are we going to do? It's
429. 15:29still important to think about what
430. 15:30model you need for the task. Now, for me
431. 15:32in Claude Code, I always have Sonic 5.5
432. 15:35selected at medium effort level. That's
433. 15:38kind of a baseline default for me for
434. 15:40most tasks. If I want to bump that up,
435. 15:42I'm going to use like inside of Hermes,
436. 15:44I'm probably going to use a Chimera K3.
437. 15:46Maybe I'll use a GPT-6 Astra and I still
438. 15:50keep it on medium. Now, if I'm trying to
439. 15:53build out a an entire project for a
440. 15:56client, like right now I'm building out
441. 15:57this tool for an exam sensor in the UK,
442. 16:01for that I'm using GPT-6 Astra on extra
443. 16:03high. I very rarely go to max or ultra,
444. 16:07but those are options as well. It's
445. 16:09worth checking out what the differences
446. 16:10are with these different effort levels.
447. 16:12And so, what I mean by effort is the the
448. 16:15thinking time. So, how much thinking is
449. 16:17going to go on behind the scenes before
450. 16:19you receive your response. So, if you're
451. 16:21asking for something short and simple,
452. 16:23you don't want to have a high effort
453. 16:25level. It's going to use more tokens,
454. 16:27it's going to take longer, and you're
455. 16:28just going to get exactly the same
456. 16:30response back. So, if we come in to
457. 16:32Hermes here, we can say, "Hey, can you
458. 16:34add a rule in the agents.md file to
459. 16:37specify that all sub agents should use
460. 16:40the the fastest, cheapest model
461. 16:43available." So, for example, Haiku 5.5
462. 16:45is fast and it's cheap, it's great for
463. 16:47sub agents. So, that's something you can
464. 16:49try out. You can add it as a rule. So,
465. 16:51every time you spin up a sub agent, it's
466. 16:53going to use a specific model. Now,
467. 16:55Claude Code does this kind of
468. 16:56automatically at the moment, but it is
469. 16:58worth mentioning in your agents.md file.
470. 17:01So, that's pretty much everything. If
471. 17:03you want to know how to see your skills
472. 17:05and MCPs, you can just ask, "Hey, can
473. 17:07you show me a list of MCPs and skills
474. 17:08that I've got installed here for
475. 17:10Hermes?" He's going to show you a list,
476. 17:11and then you can think, "Well, I've I've
477. 17:13never used this, I never never used
478. 17:14this, I never used this." Again, just by
479. 17:16asking, "Hey, can you remove these
480. 17:18skills, remove these MCPs?" And you're
481. 17:20going to be good to go. And after
482. 17:21applying all of these changes that I've
483. 17:23outlined in this video, you should get
484. 17:25some huge savings. Now, will you get 95%
485. 17:28savings? I don't know. If you've never
486. 17:30audited how you're using your AI agent,
487. 17:34then there's probably going to be quite
488. 17:35a lot of bloat, especially if you've
489. 17:37been adding rules, you've been adding
490. 17:39MCPs, you've been adding skills, you've
491. 17:41got a ton of MD files that get loaded on
492. 17:43that first prompt. I can guarantee you
493. 17:45you're going to be saving a huge amount
494. 17:47of tokens. So, we can see here five MCP
495. 17:50servers configured and 184 skills. So,
496. 17:54we can disable some of these. You can
497. 17:55see I've already disabled 63 of them.
498. 17:58So, you can read through and see what it
499. 18:00is that you need and what you don't
500. 18:01need. So, one thing real quick, if you
501. 18:03know AI could save your business time
502. 18:05and money, but you really have no idea
503. 18:07what to automate first, you can book a
504. 18:10free 30-minute AI audit call with me.
505. 18:12The link is below in the description.
506. 18:14We're going to look at your current
507. 18:15processes and your bottlenecks to find
508. 18:17the repetitive work that AI could take
509. 18:20off your team's plate. Then we'll choose
510. 18:21the single highest ROI automation to
511. 18:24build first. And then we map out how to
512. 18:26get it built, whether I build it for you
513. 18:28or whether we do it together one-on-one
514. 18:29on my consulting calls. And the link is
515. 18:32below this video. It's in the
516. 18:33description. And if you want to have
517. 18:34that conversation, you're going to speak
518. 18:36directly to me. I'm looking forward to
519. 18:38it. So, that's the whole setup I use to
520. 18:39cut my Homies cost by up to 95% without
521. 18:42sacrificing the quality of my work. So,
522. 18:44for the 24/7 Homies setup, you can use
523. 18:47hosting a VPS. The link is in the
524. 18:49description below with the code Growth
525. 18:51Lab 10, you you're going to get an extra
526. 18:5310% off. Anyway, thanks for watching.
527. 18:55I'll see you in the next one.