YouTube2Text

Tek CEO, 4 Ajan: Claude Code'da Token Testi — Transcript

by Selma Kocabıyık · 2,014 words · 299 segments · language en · Watch on YouTube

Full transcript

  1. 0:00When I ran it from the wrong folder, the
  2. 0:02agent spent about 1.5 times more tokens
  3. 0:04.
  4. 0:05But when I gave each agent the right
  5. 0:07model in the correct folder, the same
  6. 0:08package cost three times less.
  7. 0:10You really liked the office logic.
  8. 0:12That's why I came straight to the second
  9. 0:13part.
  10. 0:14In the first part, we set up an office
  11. 0:17with a few agents inside, and I mostly
  12. 0:19showed you the interface and mentioned
  13. 0:21that the agents were connected to a CEO
  14. 0:23.
  15. 0:25Now, in this episode, I will explain how
  16. 0:27much the architecture of the agent
  17. 0:28system we connected in the background
  18. 0:30spends on tokens, where you can use it,
  19. 0:32how you can use it, and how you can set
  20. 0:34up that agent architecture most
  21. 0:35efficiently in the background.
  22. 0:38Now, first, let's look at other examples
  23. 0:40. Because this idea isn't actually new.
  24. 0:42I mean, we are in a period where it's
  25. 0:44already circulating on social media and
  26. 0:45people are starting to build their own
  27. 0:47agent architectures.
  28. 0:49For example, there is an application
  29. 0:51that instantly turns a new terminal or a
  30. 0:53new chat you open into a character in an
  31. 0:56office.
  32. 0:58One even has a manager who distributes
  33. 1:00tasks for you.
  34. 1:01It is essentially the exact equivalent
  35. 1:03of the CEO I set up in my own company.
  36. 1:05All these applications have one thing in
  37. 1:08common. And that is, you can design the
  38. 1:10interface however you want. If you have
  39. 1:12a single project file, then you don't
  40. 1:14need an agent office. But you are
  41. 1:15working on many projects. You are doing
  42. 1:18big things.
  43. 1:19With that said, the lesson we should
  44. 1:20take from this is not the interface, but
  45. 1:22the way agents are connected to each
  46. 1:23other—the architecture.
  47. 1:25Now, let's imagine we have 15 open
  48. 1:26projects in front of us.
  49. 1:28When you start a chat in a project
  50. 1:30folder, there is an agent assigned to it
  51. 1:31.
  52. 1:33Now, would you want to go to 15
  53. 1:35terminals one by one or open 15 chat
  54. 1:37windows and give them different tasks
  55. 1:39one by one?
  56. 1:41No. It's too complicated. Moreover, in
  57. 1:43this logic, the agent is only
  58. 1:45responsible for its own project folder.
  59. 1:48If I go to the agent in that project and
  60. 1:50say I want to do this for another
  61. 1:51project, first, that agent's context—
  62. 1:53that is, the chat window—gets cluttered
  63. 1:55.
  64. 1:56Second, the agent will say, "I have to
  65. 1:57go to that file now, research it, look
  66. 1:59through it, and then I have to give you
  67. 2:01information." However, its Claude MD is
  68. 2:03within that project file. It would do it
  69. 2:06, not that it wouldn't, but would it be
  70. 2:08a clean way of working? No, and this
  71. 2:10could also cause you to spend more
  72. 2:12tokens. That is why we keep all projects
  73. 2:14separate. Now let's go back to my small
  74. 2:17company. What do we see here? We see a
  75. 2:19Reels machine. For example, the Reels
  76. 2:22machine has its own project. The
  77. 2:23research team has its own project. It
  78. 2:25has its own tools to use. It has its own
  79. 2:28Claude MD. It has its own way of working
  80. 2:29.
  81. 2:30If I go to the research project and tell
  82. 2:32it to get up and make Reels for me,
  83. 2:34first of all, if it doesn't go to the
  84. 2:36Reels machine and read the GitHub repos
  85. 2:38, skills, all the ability files I wrote
  86. 2:40, and all the instructions I gave inside
  87. 2:42the Reels machine, it won't produce the
  88. 2:44result I want.
  89. 2:45It's like going to your accountant and
  90. 2:46saying, "I'm moving you to the
  91. 2:47advertising department now." It's
  92. 2:48essentially like telling them to go to
  93. 2:50the advertising department and start
  94. 2:51making ads. It was a problem I needed to
  95. 2:53solve. What did I do? That's why I added
  96. 2:55a CEO to my company. Look, there might
  97. 2:57be 10 different ways to solve this
  98. 2:58problem. You can mention them in the
  99. 3:00comments, but this was my way of solving
  100. 3:01it. So, how does the CEO make my job
  101. 3:03easier? For instance, bosses don't go to
  102. 3:05accounting and give tasks directly to
  103. 3:06the accountant. They don't talk to them
  104. 3:08. Under normal circumstances, a boss has
  105. 3:10an assistant. They explain their needs
  106. 3:11to the assistant. And the assistant
  107. 3:12visits all the rooms in the office. They
  108. 3:14solve the problem. At this point, I go
  109. 3:16to my CEO and say, "I'll be working on
  110. 3:18this today.""Set it up." And it takes
  111. 3:20that agent to its own project folder. It
  112. 3:23just brings the finished work and the
  113. 3:24final result to my desk. That's the
  114. 3:26whole point. Now, let's get to the
  115. 3:28sub-agent logic. Who calls the sub-agent
  116. 3:29, then? Only my agents call the
  117. 3:31sub-agents. For example, while the Reels
  118. 3:33machine is preparing Reels, it calls 4,
  119. 3:345, or 6 sub-agents. And after the job is
  120. 3:36finished, it shuts them down. We create
  121. 3:37permanent agents that wait at the head
  122. 3:39of every project folder. These agents
  123. 3:41keep the accumulation of weeks and
  124. 3:43months in that folder. This is actually
  125. 3:45why it's great for agents to talk to
  126. 3:47each other in the background. Because,
  127. 3:50for example, let's say I'm doing
  128. 3:51research within a project for my own
  129. 3:53video. And I'm converting it into a
  130. 3:55YouTube video within the chat. So, think
  131. 3:56of one as dealing with editing, and the
  132. 3:58other as dealing with research. Under
  133. 4:00normal circumstances, we could carry the
  134. 4:02output of one to the other and say, "
  135. 4:03These are the research results.""Add
  136. 4:05this to the YouTube video." But at this
  137. 4:07point, I can now say, "There is a
  138. 4:09problem like this.""The research team
  139. 4:11and your YouTube video editing side need
  140. 4:13to talk to each other and solve this
  141. 4:15problem."
  142. 4:15Here, I have already provided the agents
  143. 4:17'memory with the fact that they need to
  144. 4:19work together. This could be a
  145. 4:21capability file for it.
  146. 4:23It could be a line I wrote in the system
  147. 4:25prompt, or it could be information we
  148. 4:27saved to their memory before.
  149. 4:29Here, the agent knows they need to talk
  150. 4:31and solve my problem together. That's
  151. 4:33what I mean by that agent architecture
  152. 4:35in the background. Now, let's get to the
  153. 4:37token and cost issue. Is it better to go
  154. 4:40and give a single prompt at once, or to
  155. 4:42explain what you want?
  156. 4:44Which one burns more tokens and what is
  157. 4:46the difference between them? Now, here
  158. 4:48comes our honest answer. It probably
  159. 4:50doesn't make much difference in terms of
  160. 4:52tokens. What do I mean? The difference
  161. 4:54here is making the way of working as
  162. 4:56efficient as possible. The chats stay
  163. 4:58clean. Tasks don't get mixed up. You
  164. 5:00don't explain the same thing twice.
  165. 5:02Everyone knows their own job. We
  166. 5:04conducted a test to measure whether this
  167. 5:07system actually works more efficiently.
  168. 5:10And in this test, we assigned five
  169. 5:11different tasks. We had the same
  170. 5:13publishing package—the YouTube
  171. 5:15description, the Reels script, and the
  172. 5:17Instagram and X text—done in two ways.
  173. 5:20In one, we opened the terminal inside a
  174. 5:22completely different project and asked
  175. 5:24it to complete the task in a different
  176. 5:25project file.
  177. 5:27In the other, we delivered the same job
  178. 5:29to the correct agent within its own
  179. 5:31project file, and when done from the
  180. 5:33wrong folder, the agent spent about 1.5
  181. 5:35times more tokens. Because it had to
  182. 5:37find the right project first and read
  183. 5:39its rules. But by providing the right
  184. 5:42model for each agent in the correct
  185. 5:44folder, the same package became three
  186. 5:45times cheaper, and the cost of the CEO
  187. 5:47agent distributing the work was about 1
  188. 5:50cent per task.
  189. 5:51So, we can see that this system actually
  190. 5:54performs better for you both in terms of
  191. 5:57tokens and efficiency.
  192. 5:59Now, normally, the specific tasks Cloud
  193. 6:01Code does in the background is a model
  194. 6:03that works great on factory settings.
  195. 6:06But what we are doing here is guiding
  196. 6:08the Cloud agent at a certain point to
  197. 6:10both increase its efficiency and reduce
  198. 6:13token usage as much as possible. Now,
  199. 6:15the real cost trap is somewhere else
  200. 6:17entirely. And it is this. Some people
  201. 6:19really go ahead, set up devices, and run
  202. 6:22hundreds of agents. Look, this really
  203. 6:25spends tokens. Because you might not
  204. 6:27need to run 50 agents at the same time.
  205. 6:29You know I have many different videos on
  206. 6:31agents and systems, especially on AI
  207. 6:32automations. For example, I set up
  208. 6:34agents that work with a timer, and these
  209. 6:36are agents that spring into action when
  210. 6:38their time comes. In this office logic,
  211. 6:41all agents are sleeping, but you didn't
  212. 6:42see the working Hermes agent in the back
  213. 6:44, and these live in the cloud. Even if
  214. 6:46my computer is turned off, these guys
  215. 6:48keep on working.
  216. 6:50Look, they are listening 24/7, but they
  217. 6:52don't spend tokens while listening. The
  218. 6:55agents inside this office are completely
  219. 6:56shut down and sleeping. They only start
  220. 6:58working when I send something to the CEO
  221. 7:00. And the bad part is, since they don't
  222. 7:02live in the cloud, they don't continue
  223. 7:04working when the computer is off. Now,
  224. 7:06I'm coming to those who want to set this
  225. 7:07up from scratch. Let's sketch out an
  226. 7:09example scenario right here. Okay. The
  227. 7:11first thing we do is open a project
  228. 7:12folder. This will be our accounting
  229. 7:14agent.
  230. 7:15You upload all the skill files you've
  231. 7:17found related to accounting, your own
  232. 7:19data, and important information here.
  233. 7:21As much data as there is related to
  234. 7:23accounting goes into this folder. It
  235. 7:24will stay clean. Then our second folder
  236. 7:26is Gmail. Our 3rd folder is marketing.
  237. 7:27Our 4th. folder is research. Look, 4
  238. 7:29agents are already ready. But the
  239. 7:31important thing here is, don't just open
  240. 7:33Gmail and leave it empty. Fill it up.
  241. 7:35What do you want to be done inside Gmail
  242. 7:36? Place all the information inside.
  243. 7:38Which Gmail accounts will be connected?
  244. 7:40What drafts will be used when replying
  245. 7:41to Gmails? By the way, we will spend
  246. 7:43tokens at the beginning. Because whether
  247. 7:44you want to or not, you spend tokens
  248. 7:46while creating the structure of the
  249. 7:47folders. Later on, it starts to become
  250. 7:49efficient. What are we doing now? We are
  251. 7:51opening one more folder. Its name will
  252. 7:53be the CEO folder. We will write where
  253. 7:55our four agents are and what each one
  254. 7:58does. And you won't need to talk to your
  255. 8:00four agents or four folders one by one
  256. 8:02anymore. Because the CEO will handle it
  257. 8:05for you. If you are happy with the chat
  258. 8:07window, that's fine. You can continue
  259. 8:09like that, but this is actually the
  260. 8:10essence and logic of the agent company
  261. 8:11that has been very popular in the last
  262. 8:13month. Now, the difference you see in
  263. 8:15this test doesn't just come from the
  264. 8:16folders being separate. It is also very
  265. 8:18important what is inside each folder.
  266. 8:20For example, how long your rule file is
  267. 8:22, how you keep the memory, what the
  268. 8:24agent reads at every startup; for
  269. 8:26instance, the machine that spent the
  270. 8:28most tokens in the test was the one with
  271. 8:30the longest rule file. So these directly
  272. 8:32affect both your token usage and
  273. 8:34efficiency. That is why you need to set
  274. 8:36up your project folders according to
  275. 8:38your own workflow. Of course, at this
  276. 8:40point, you can determine the size of
  277. 8:42your own rules. However, it is always
  278. 8:44necessary to keep them as short and
  279. 8:45concise as possible. Now, of course,
  280. 8:47since the release of Cloud Code, such
  281. 8:49topics, especially project structure and
  282. 8:52how to set up an agent system, have been
  283. 8:54discussed for months.
  284. 8:56But if you are curious about the filing
  285. 8:58architecture within a project from an
  286. 9:01office perspective, perhaps we can make
  287. 9:03another episode on that.
  288. 9:05Let me know your thoughts in the
  289. 9:06comments. By the way, the token amounts
  290. 9:08you saw in the previous video were what
  291. 9:10my agents running in the background
  292. 9:12spent. That wasn't the amount spent
  293. 9:14because of the system we set up.
  294. 9:16Also, the screenshots were from
  295. 9:17different times, so it's normal that the
  296. 9:19tokens you saw there don't match their
  297. 9:21prices; you can find the interface in
  298. 9:22the git repo.
  299. 9:24Good luck. See you in the next video.

About this transcript

This page contains the full transcript of Tek CEO, 4 Ajan: Claude Code'da Token Testi by Selma Kocabıyık, generated from the public captions YouTube serves with the video. The transcript has 2,014 words across 299 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.