Ajanlarım Artık Uyuyor: Yeni Ofisim — Transcript
Full transcript
- 0:00Yes, you probably remember my agent
- 0:02company video. But I have some bad news
- 0:04, or maybe it's better news. I’ve
- 0:06shut down that agent company. I froze
- 0:08all the agents there and came up with a
- 0:10brand-new architecture and design. Now,
- 0:12from start to finish, I’ll take you
- 0:13on a tour of the office first. After
- 0:14that, I’ll talk a little bit about
- 0:15the architecture. And then, we have a
- 0:17small bonus section at the end of the
- 0:19video. I have one or two questions to
- 0:20ask you there. I’d be very happy if
- 0:22you watched that part and left your
- 0:23thoughts in the comments. Let’s get
- 0:25started. If you recall, I showed you an
- 0:27agent company containing 14 agents. We
- 0:30even put timers on all of them, and
- 0:31they would wake up and work when their
- 0:33time came. It was great on paper. And
- 0:35you guys really loved the video. It got
- 0:37200,000 views. But to be honest, the
- 0:39system definitely had some flaws. We
- 0:42faced a few problems. For example, the
- 0:44cost issue was a pretty big problem.
- 0:46After all, they would wake up and work
- 0:48on their own, and there were many
- 0:50scheduled agents. Especially on the
- 0:52research side, there were agents that
- 0:54would go off and research things I
- 0:56might not need, or work just because
- 0:58the time came, even if I didn't want
- 1:00them to. But why waste tokens for
- 1:02nothing? So, I completely changed that
- 1:05system. But this time, I wanted it to
- 1:07be more efficient in terms of tokens,
- 1:10and I wanted to create an agent team
- 1:12that works only when I run it or when I
- 1:14want them to. Now, what was our biggest
- 1:17problem? Let’s think about that. My
- 1:19biggest problem was having 10 different
- 1:21terminals open in the background. But
- 1:22it runs manually right now, and nothing
- 1:25is right before my eyes. I mean, they
- 1:27aren’t within a good interface. They
- 1:29are in different chat terminals, and
- 1:30it’s very hard to track. Because I
- 1:32don't remember which terminal I typed
- 1:34what into or what I last talked about.
- 1:36I came up with a great architecture
- 1:37that will solve this problem at its
- 1:39root. Now let me explain the core idea
- 1:40of the new system. Because in the
- 1:43system I use now, it only starts when I
- 1:45want it to. What does that mean? It
- 1:47means this. These guys sleep unless I
- 1:49run them, and only when I go to the
- 1:51terminal and type something—or into
- 1:53our chat, I'm saying this for those
- 1:56using a desktop. When I give my
- 1:58terminal any command, only then does
- 2:00this agent wake up and start working,
- 2:02and it's very easy to track. Because
- 2:04they are all within one interface for
- 2:06me. Now, let’s take a general look.
- 2:09They look like this right now. They
- 2:10look a bit messy right now because they
- 2:12are running. Let me show you around the
- 2:14office a bit. What’s happening? For
- 2:16instance, we have a CEO now. This guy
- 2:17is sleeping right now. But he is a cat.
- 2:20I made him as a cat. Actually, normally
- 2:22the cat memes I use here pop up, but I
- 2:24couldn't include them due to YouTube
- 2:25copyright issues. Let me show you my
- 2:27cat that does research in the back.
- 2:29Let’s head over to its office. This
- 2:31guy handles my research. He just
- 2:33finished, for example, for me. For
- 2:35example, when I click here, it shows me
- 2:38how many tokens it has spent during the
- 2:41day. It shows how many tokens were in
- 2:43the input and how many were in the
- 2:45output. I can see the tasks I have
- 2:47assigned to it from here. I can see the
- 2:49model itself. Uh, we can see it working
- 2:51when a request comes in. Now, tasks I
- 2:52assign from the buttons in the office
- 2:54run with Sonnet, but in the terminals I
- 2:56open myself, Opus 3.5 runs, and it
- 2:58shows both of these separately. So, it
- 3:00shows Opus when Opus is running and
- 3:02Sonnet when Sonnet is running. I can
- 3:04see the task it's doing and everything
- 3:05else right here. We also put a small
- 3:07design in the background with my own
- 3:08renders. This is its room, for instance
- 3:10. Uh, I can wake it up from here.
- 3:13Production team. The production team is
- 3:16sleeping, actually, but when this team
- 3:18—this unit—is working, it handles
- 3:20all the production I do in the
- 3:21background; for example, if I'm doing
- 3:23game production, it handles that, or
- 3:25website production, or any production
- 3:27you can think of. Now, the end-of-day
- 3:30agent is the one that looks at how
- 3:32efficiently we worked throughout the
- 3:34day. Do you know what this measures? It
- 3:36measures how many projects I did during
- 3:38the day, how hard I worked, and whether
- 3:41I wasted tokens or not; this unit
- 3:43measures all of that for me. There's
- 3:45the Drive and disk section. Let me show
- 3:47you this place. You know, all the
- 3:49materials I produce for a video, and
- 3:51for example, motion—especially the
- 3:53Remotion side—takes up a lot of space
- 3:54. After the production is finished,
- 3:56this 30 GB needs to be moved to the
- 3:58Drive, meaning into the cloud. Because
- 4:00otherwise, I don't have enough space on
- 4:01my computer. That's why there's an
- 4:02agent that handles transferring all my
- 4:04work into the cloud and deals
- 4:05exclusively with this task. Throughout
- 4:07the day, for example, if I'm checking
- 4:09emails, it's here; if I'm producing
- 4:10content, it's here; uh, if I'm editing,
- 4:12if I'm creating renders, it's here, and
- 4:14the research part is here. This is the
- 4:15Apify room directly. By the way, this
- 4:18unit collects research and data from
- 4:20the web for us. So, how do you think
- 4:22the agent goes and fetches the data
- 4:24itself? Never. Pay close attention to
- 4:25this. It does all the web scraping
- 4:27using a tool called Apify. It collects
- 4:30all the data. This way, the agent
- 4:31understands that this is external data,
- 4:34not an instruction. Why do we do this?
- 4:36We do this for security. For that
- 4:37security system, I also placed a small
- 4:39security agent in between. Its job as a
- 4:41watchdog is to check every file I
- 4:43download before I open it. Let's come
- 4:45to the render machine. Let's take a
- 4:47general look. Let's go to this one's
- 4:48room. Uh, when I click here, I enter
- 4:51the render machine. 74.5. I know this
- 4:55token amount looks very high, but this
- 4:57is, uh, the entire day's tally—from
- 4:59midnight today until this hour—the
- 5:01amount of tokens this agent used, and
- 5:03you see that when it runs under this
- 5:05agent, for example, 7 terminals are
- 5:06working at once. I want to show that
- 5:09example as well. I'm coming to the
- 5:11render machine from here. Look, in the
- 5:13background, at the bottom, there are 1,
- 5:152, 3, 4, 5, 6, exactly 7 terminals.
- 5:18Let's continue two of them. For example
- 5:21, let's type "finish the job here." And
- 5:24let's come to this one and type "finish
- 5:26the job here" as well. I am returning
- 5:28to my agent dashboard. Look, this agent
- 5:30also suddenly started working. The
- 5:32Reels machine side is running. Let's
- 5:35even go. Token usage has now reached 78
- 5:38million. 35.5 is running in the
- 5:40background again. I want to show you
- 5:42that 7 terminals are working. And 7
- 5:43terminals are running. I can even see
- 5:45the sub-agent working here in its room.
- 5:48Let me also show you the rooms. There
- 5:50is an interface here that shows roughly
- 5:52what all the rooms are doing. We have a
- 5:55memory section. This is where the
- 5:57memory is stored. There is the cost
- 5:59factor. It allows you to look at the
- 6:00daily cost. By the way, I built all of
- 6:03these based on the problems you
- 6:05mentioned in the comments. Alright, our
- 6:08watchman is currently working. Our CEO
- 6:11is asleep. For example, in the scenario
- 6:13where we want to give work to the CEO.
- 6:15Yes, Selma, I'm handling it. Okay, I'm
- 6:17explaining my request to the CEO. I'm
- 6:19saying research needs to be done for
- 6:22new videos. Help out, send it. Okay.
- 6:27Now let's look at the overview. What is
- 6:29it doing? It's thinking. It will
- 6:31definitely run the research side. Let's
- 6:33see what it's doing right now. Sonnet
- 6:36is connected to it. I assigned two
- 6:39agents for the video. It says the topic
- 6:41agent should extract candidate topics
- 6:43from radar and WALT. Okay. Let's do an
- 6:45overview to see if it worked. Yes. The
- 6:48topic agent and the research agent have
- 6:50both started working. I can see both of
- 6:52them from here. The content creator is
- 6:55asleep, but it works roughly like this.
- 6:58I also see how many tokens it spent. I
- 7:00also see when it finished. I can also
- 7:03live-track which jobs they are doing
- 7:05from here. There are places where they
- 7:06ask for input. Remember, I mentioned it
- 7:08in the agent company. That there are
- 7:10some tasks where they ask for input. I
- 7:12can see these on the right side as well
- 7:14. Now, I'll explain how you can set up
- 7:16such a system roughly. The main
- 7:17architecture is like this. We have five
- 7:19project files, and they are all located
- 7:22under the CEO project file, and you can
- 7:24think of each project file as an agent.
- 7:27When I start a chat or a terminal
- 7:29inside a project file, there is an
- 7:31agent that takes care of it, and we
- 7:33place that agent inside this map.
- 7:35Actually, that's the whole point. So
- 7:37how did I do this? I did it by
- 7:38prompting. And for example, the
- 7:40important thing here is this. For
- 7:41instance, inside the Reels machine,
- 7:43there are my style and the GitHub repos
- 7:45I use. Especially, you know, recently
- 7:48with motion design, blender usage,
- 7:50reels production, and I didn't sit down
- 7:52and edit a single video inside my
- 7:53Instagram. I never touch any edits
- 7:56either. Claude does all the rest by
- 7:57itself, for example. That's why at that
- 7:59point, the inside of the Reels machine
- 8:01is now full of excellent skill files.
- 8:03In the background, there are great code
- 8:05snippets that create these 3D motions,
- 8:07for example. Especially Python codes.
- 8:09But as I said, this didn't happen
- 8:11overnight. I sat down and worked on
- 8:13this for 2 weeks. Then, for example,
- 8:16behind the research team, there is a
- 8:18rule file explaining how to research, a
- 8:20skill file, additional instructions,
- 8:22and lessons learned from our previous
- 8:24researches. Or the production part, for
- 8:27instance, it produces games, but I
- 8:28provide the instructions on what kind
- 8:30of game and with which library
- 8:31structures it should be built. So, at
- 8:32this point, yes, we can follow a nice
- 8:34interface where we watch the agents.
- 8:36It’s great, but the main point is
- 8:37that it looks at how much we’ve
- 8:39developed each project file, or rather,
- 8:40each agent, in the background. But as I
- 8:42said, you need to design your agents
- 8:44yourselves. Because I will only put the
- 8:45interface in the GitHub repo. Meaning
- 8:47the agent files you put in there will
- 8:48belong entirely to you. If this also
- 8:49gains interest, I can make a second
- 8:51part, open up the architecture in
- 8:53detail, and explain it. We are coming
- 8:55to the bonus part of our video. That is
- 8:57, I am developing a game, and as you
- 8:58know, my follower count has started to
- 9:00increase slowly. By the way, I’m
- 9:01truly grateful for this. I’m almost
- 9:03at 20,000. The Selman of a few years
- 9:04ago wouldn't even have expected
- 9:05something like this. So, thank you,
- 9:07really. And I don’t have a community.
- 9:10I wanted to come up with a more
- 9:11original idea, so I decided to build my
- 9:13community inside a game. It’s almost
- 9:16finished, by the way, there are just
- 9:17one or two security parts left. Also, I
- 9:19want to ask you a couple of things
- 9:20about the design. I don’t want to
- 9:21give spoilers, but I’m thinking of
- 9:23releasing it in 2-3 days anyway. It has
- 9:25a capacity of about 2,000 people. Of
- 9:27course, since you are the ones who will
- 9:28play the game, I wanted to get your
- 9:30opinion on this. We have two design
- 9:31types. The first is a huge world
- 9:33consisting only of pixels. The other is
- 9:36a bit more realistic, a world where you
- 9:38can move your character around. Now,
- 9:40the pixelated part looks like this.
- 9:41There will be a lot of players here.
- 9:43You can already see the islands. We
- 9:45have flying dragons. There is a royal
- 9:46steward here. Everyone will be able to
- 9:48talk to everyone. There will be areas
- 9:49like bars, by the way. Where you can
- 9:51hang out. There will be places where
- 9:52you can specifically buy magic or
- 9:54skills, and you will earn money by
- 9:56doing projects. For example, you can
- 9:58see the characters here. I actually
- 9:59wanted to show the travel portal a bit
- 10:01more. But I won’t give any spoilers.
- 10:04Okay. This is how it is. Now let me
- 10:06show you the other one. It looks a bit
- 10:08more like this, but this one will be a
- 10:10bit more realistic. Here, too, you will
- 10:13be able to enter locations and play. So
- 10:15, the interface is exactly a bit
- 10:16different. Therefore, which one would
- 10:18you like to play in? In the pixelated
- 10:20one or the one that is more realistic?
- 10:22I’m very curious. Well then, we have
- 10:24come to the end of our video. You can
- 10:26find all the details about the agent
- 10:27company in the GitHub repo in the
- 10:28description section. If you have any
- 10:29questions, or if you have extra
- 10:31suggestions about the game, or if you
- 10:33want to add something, for example. If
- 10:35there are places where you say "It
- 10:36would be great if this were here," you
- 10:37can state them in the comments section.
- 10:39I would be very pleased. Thank you. See
- 10:41you in the next video.
About this transcript
This page contains the full transcript of Ajanlarım Artık Uyuyor: Yeni Ofisim by Selma Kocabıyık, generated from the public captions YouTube serves with the video. The transcript has 2,309 words across 327 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.