Tek CEO, 4 Ajan: Claude Code'da Token Testi — Transcript
Full transcript
- 0:00When I ran it from the wrong folder, the
- 0:02agent spent about 1.5 times more tokens
- 0:04.
- 0:05But when I gave each agent the right
- 0:07model in the correct folder, the same
- 0:08package cost three times less.
- 0:10You really liked the office logic.
- 0:12That's why I came straight to the second
- 0:13part.
- 0:14In the first part, we set up an office
- 0:17with a few agents inside, and I mostly
- 0:19showed you the interface and mentioned
- 0:21that the agents were connected to a CEO
- 0:23.
- 0:25Now, in this episode, I will explain how
- 0:27much the architecture of the agent
- 0:28system we connected in the background
- 0:30spends on tokens, where you can use it,
- 0:32how you can use it, and how you can set
- 0:34up that agent architecture most
- 0:35efficiently in the background.
- 0:38Now, first, let's look at other examples
- 0:40. Because this idea isn't actually new.
- 0:42I mean, we are in a period where it's
- 0:44already circulating on social media and
- 0:45people are starting to build their own
- 0:47agent architectures.
- 0:49For example, there is an application
- 0:51that instantly turns a new terminal or a
- 0:53new chat you open into a character in an
- 0:56office.
- 0:58One even has a manager who distributes
- 1:00tasks for you.
- 1:01It is essentially the exact equivalent
- 1:03of the CEO I set up in my own company.
- 1:05All these applications have one thing in
- 1:08common. And that is, you can design the
- 1:10interface however you want. If you have
- 1:12a single project file, then you don't
- 1:14need an agent office. But you are
- 1:15working on many projects. You are doing
- 1:18big things.
- 1:19With that said, the lesson we should
- 1:20take from this is not the interface, but
- 1:22the way agents are connected to each
- 1:23other—the architecture.
- 1:25Now, let's imagine we have 15 open
- 1:26projects in front of us.
- 1:28When you start a chat in a project
- 1:30folder, there is an agent assigned to it
- 1:31.
- 1:33Now, would you want to go to 15
- 1:35terminals one by one or open 15 chat
- 1:37windows and give them different tasks
- 1:39one by one?
- 1:41No. It's too complicated. Moreover, in
- 1:43this logic, the agent is only
- 1:45responsible for its own project folder.
- 1:48If I go to the agent in that project and
- 1:50say I want to do this for another
- 1:51project, first, that agent's context—
- 1:53that is, the chat window—gets cluttered
- 1:55.
- 1:56Second, the agent will say, "I have to
- 1:57go to that file now, research it, look
- 1:59through it, and then I have to give you
- 2:01information." However, its Claude MD is
- 2:03within that project file. It would do it
- 2:06, not that it wouldn't, but would it be
- 2:08a clean way of working? No, and this
- 2:10could also cause you to spend more
- 2:12tokens. That is why we keep all projects
- 2:14separate. Now let's go back to my small
- 2:17company. What do we see here? We see a
- 2:19Reels machine. For example, the Reels
- 2:22machine has its own project. The
- 2:23research team has its own project. It
- 2:25has its own tools to use. It has its own
- 2:28Claude MD. It has its own way of working
- 2:29.
- 2:30If I go to the research project and tell
- 2:32it to get up and make Reels for me,
- 2:34first of all, if it doesn't go to the
- 2:36Reels machine and read the GitHub repos
- 2:38, skills, all the ability files I wrote
- 2:40, and all the instructions I gave inside
- 2:42the Reels machine, it won't produce the
- 2:44result I want.
- 2:45It's like going to your accountant and
- 2:46saying, "I'm moving you to the
- 2:47advertising department now." It's
- 2:48essentially like telling them to go to
- 2:50the advertising department and start
- 2:51making ads. It was a problem I needed to
- 2:53solve. What did I do? That's why I added
- 2:55a CEO to my company. Look, there might
- 2:57be 10 different ways to solve this
- 2:58problem. You can mention them in the
- 3:00comments, but this was my way of solving
- 3:01it. So, how does the CEO make my job
- 3:03easier? For instance, bosses don't go to
- 3:05accounting and give tasks directly to
- 3:06the accountant. They don't talk to them
- 3:08. Under normal circumstances, a boss has
- 3:10an assistant. They explain their needs
- 3:11to the assistant. And the assistant
- 3:12visits all the rooms in the office. They
- 3:14solve the problem. At this point, I go
- 3:16to my CEO and say, "I'll be working on
- 3:18this today.""Set it up." And it takes
- 3:20that agent to its own project folder. It
- 3:23just brings the finished work and the
- 3:24final result to my desk. That's the
- 3:26whole point. Now, let's get to the
- 3:28sub-agent logic. Who calls the sub-agent
- 3:29, then? Only my agents call the
- 3:31sub-agents. For example, while the Reels
- 3:33machine is preparing Reels, it calls 4,
- 3:345, or 6 sub-agents. And after the job is
- 3:36finished, it shuts them down. We create
- 3:37permanent agents that wait at the head
- 3:39of every project folder. These agents
- 3:41keep the accumulation of weeks and
- 3:43months in that folder. This is actually
- 3:45why it's great for agents to talk to
- 3:47each other in the background. Because,
- 3:50for example, let's say I'm doing
- 3:51research within a project for my own
- 3:53video. And I'm converting it into a
- 3:55YouTube video within the chat. So, think
- 3:56of one as dealing with editing, and the
- 3:58other as dealing with research. Under
- 4:00normal circumstances, we could carry the
- 4:02output of one to the other and say, "
- 4:03These are the research results.""Add
- 4:05this to the YouTube video." But at this
- 4:07point, I can now say, "There is a
- 4:09problem like this.""The research team
- 4:11and your YouTube video editing side need
- 4:13to talk to each other and solve this
- 4:15problem."
- 4:15Here, I have already provided the agents
- 4:17'memory with the fact that they need to
- 4:19work together. This could be a
- 4:21capability file for it.
- 4:23It could be a line I wrote in the system
- 4:25prompt, or it could be information we
- 4:27saved to their memory before.
- 4:29Here, the agent knows they need to talk
- 4:31and solve my problem together. That's
- 4:33what I mean by that agent architecture
- 4:35in the background. Now, let's get to the
- 4:37token and cost issue. Is it better to go
- 4:40and give a single prompt at once, or to
- 4:42explain what you want?
- 4:44Which one burns more tokens and what is
- 4:46the difference between them? Now, here
- 4:48comes our honest answer. It probably
- 4:50doesn't make much difference in terms of
- 4:52tokens. What do I mean? The difference
- 4:54here is making the way of working as
- 4:56efficient as possible. The chats stay
- 4:58clean. Tasks don't get mixed up. You
- 5:00don't explain the same thing twice.
- 5:02Everyone knows their own job. We
- 5:04conducted a test to measure whether this
- 5:07system actually works more efficiently.
- 5:10And in this test, we assigned five
- 5:11different tasks. We had the same
- 5:13publishing package—the YouTube
- 5:15description, the Reels script, and the
- 5:17Instagram and X text—done in two ways.
- 5:20In one, we opened the terminal inside a
- 5:22completely different project and asked
- 5:24it to complete the task in a different
- 5:25project file.
- 5:27In the other, we delivered the same job
- 5:29to the correct agent within its own
- 5:31project file, and when done from the
- 5:33wrong folder, the agent spent about 1.5
- 5:35times more tokens. Because it had to
- 5:37find the right project first and read
- 5:39its rules. But by providing the right
- 5:42model for each agent in the correct
- 5:44folder, the same package became three
- 5:45times cheaper, and the cost of the CEO
- 5:47agent distributing the work was about 1
- 5:50cent per task.
- 5:51So, we can see that this system actually
- 5:54performs better for you both in terms of
- 5:57tokens and efficiency.
- 5:59Now, normally, the specific tasks Cloud
- 6:01Code does in the background is a model
- 6:03that works great on factory settings.
- 6:06But what we are doing here is guiding
- 6:08the Cloud agent at a certain point to
- 6:10both increase its efficiency and reduce
- 6:13token usage as much as possible. Now,
- 6:15the real cost trap is somewhere else
- 6:17entirely. And it is this. Some people
- 6:19really go ahead, set up devices, and run
- 6:22hundreds of agents. Look, this really
- 6:25spends tokens. Because you might not
- 6:27need to run 50 agents at the same time.
- 6:29You know I have many different videos on
- 6:31agents and systems, especially on AI
- 6:32automations. For example, I set up
- 6:34agents that work with a timer, and these
- 6:36are agents that spring into action when
- 6:38their time comes. In this office logic,
- 6:41all agents are sleeping, but you didn't
- 6:42see the working Hermes agent in the back
- 6:44, and these live in the cloud. Even if
- 6:46my computer is turned off, these guys
- 6:48keep on working.
- 6:50Look, they are listening 24/7, but they
- 6:52don't spend tokens while listening. The
- 6:55agents inside this office are completely
- 6:56shut down and sleeping. They only start
- 6:58working when I send something to the CEO
- 7:00. And the bad part is, since they don't
- 7:02live in the cloud, they don't continue
- 7:04working when the computer is off. Now,
- 7:06I'm coming to those who want to set this
- 7:07up from scratch. Let's sketch out an
- 7:09example scenario right here. Okay. The
- 7:11first thing we do is open a project
- 7:12folder. This will be our accounting
- 7:14agent.
- 7:15You upload all the skill files you've
- 7:17found related to accounting, your own
- 7:19data, and important information here.
- 7:21As much data as there is related to
- 7:23accounting goes into this folder. It
- 7:24will stay clean. Then our second folder
- 7:26is Gmail. Our 3rd folder is marketing.
- 7:27Our 4th. folder is research. Look, 4
- 7:29agents are already ready. But the
- 7:31important thing here is, don't just open
- 7:33Gmail and leave it empty. Fill it up.
- 7:35What do you want to be done inside Gmail
- 7:36? Place all the information inside.
- 7:38Which Gmail accounts will be connected?
- 7:40What drafts will be used when replying
- 7:41to Gmails? By the way, we will spend
- 7:43tokens at the beginning. Because whether
- 7:44you want to or not, you spend tokens
- 7:46while creating the structure of the
- 7:47folders. Later on, it starts to become
- 7:49efficient. What are we doing now? We are
- 7:51opening one more folder. Its name will
- 7:53be the CEO folder. We will write where
- 7:55our four agents are and what each one
- 7:58does. And you won't need to talk to your
- 8:00four agents or four folders one by one
- 8:02anymore. Because the CEO will handle it
- 8:05for you. If you are happy with the chat
- 8:07window, that's fine. You can continue
- 8:09like that, but this is actually the
- 8:10essence and logic of the agent company
- 8:11that has been very popular in the last
- 8:13month. Now, the difference you see in
- 8:15this test doesn't just come from the
- 8:16folders being separate. It is also very
- 8:18important what is inside each folder.
- 8:20For example, how long your rule file is
- 8:22, how you keep the memory, what the
- 8:24agent reads at every startup; for
- 8:26instance, the machine that spent the
- 8:28most tokens in the test was the one with
- 8:30the longest rule file. So these directly
- 8:32affect both your token usage and
- 8:34efficiency. That is why you need to set
- 8:36up your project folders according to
- 8:38your own workflow. Of course, at this
- 8:40point, you can determine the size of
- 8:42your own rules. However, it is always
- 8:44necessary to keep them as short and
- 8:45concise as possible. Now, of course,
- 8:47since the release of Cloud Code, such
- 8:49topics, especially project structure and
- 8:52how to set up an agent system, have been
- 8:54discussed for months.
- 8:56But if you are curious about the filing
- 8:58architecture within a project from an
- 9:01office perspective, perhaps we can make
- 9:03another episode on that.
- 9:05Let me know your thoughts in the
- 9:06comments. By the way, the token amounts
- 9:08you saw in the previous video were what
- 9:10my agents running in the background
- 9:12spent. That wasn't the amount spent
- 9:14because of the system we set up.
- 9:16Also, the screenshots were from
- 9:17different times, so it's normal that the
- 9:19tokens you saw there don't match their
- 9:21prices; you can find the interface in
- 9:22the git repo.
- 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.