Transcript of До 30 агентов одновременно! Opus 5 и GPT 5.6 Sol делают игру в Orca
Студия Игор
0:06Today, using an agentic system where an0:09 average of four agents worked0:10 simultaneously, we will try to make a0:12 game that suspiciously resembles Heroes0:15 of Might and Magic. Stay tuned, it will0:17 be interesting. Hello everyone and0:23 welcome to the Game Studio channel. My0:25 name is Igor, and today we are going to0:27 build a game. You often wrote to me in0:29 the comments: "Igor, you're making0:31 nonsense" or "You're doing everything0:34 wrong!" Who even makes games like that0:36 nowadays? Just running a prompt and0:38 making a game in one shot or one0:40 session is no longer relevant. And one0:43 comment interested me more than the0:45 others. One of you recommended that I0:48 try Orka, a so-called ADE, or Agent0:50 Development Environment. It's something0:54 very similar to a code editor, but the0:56 central focus here is interacting with0:58 the agent. There are a lot of cool1:00 features in this ADE. You can manage1:03 several projects simultaneously, run1:06 different agents in these projects, and1:08 even launch entire fleets of agents to1:11 work on one or more tasks across one or1:14 more branches. Additionally, these1:17 agents have access to the full toolset1:19 and the current ADE context. This means1:21 they can interact with each other using1:24 special commands, send messages to one1:26 another, and even forward context in1:28 certain situations. There are many,1:31 many other features there too. I won't1:33 talk about all of them right now. The1:36 project is available on GitHub and is1:38 free and open source. Everyone can try1:41 it out. I downloaded and installed Orka1:44. I didn't need any special knowledge1:46 or skills for that. There is some basic1:49 setup during the first launch. And, in1:51 general, there's no need to configure1:53 anything extra. All your agents that1:56 are already installed and running on1:57 your computer are available in Orka by2:00 default. That is, if you have, say,2:03 CodeX or some other tools, you can use2:06 them immediately in the ADE. But for me2:10, of course, I was most interested in2:13 trying out orchestration and creating a2:15 real game studio that will make a game2:17 instead of me. And for this, I prepared2:21 a skill, a working folder containing a2:24 full description of how exactly the2:26 Game Studio works. This will be needed2:30 for the producer, that is, the2:31 orchestrator, who will be directly2:33 responsible for coordinating roles for2:35 game development. And there are about2:39 10 different roles available in this2:41 project. It includes a 2D designer, a2:43 3D modeler, an architect, a developer,2:46 and so on. For each of them, a specific2:48 model is defined that should be used2:51 with that role. A prompt is also2:53 defined, specifying what this role can2:56 do and what it explicitly cannot do.2:59 The orchestrator knows how to manage3:00 all of this. When it creates an agent3:03 to implement a task, it immediately and3:06 automatically passes all these3:08 instructions for further work. By the3:12 way, this Game Studio project will be3:14 available on my GitHub, the link will3:16 be in the description, so if you want3:17 to try it, grab it from there. Next, to3:21 make a game, we need to prepare a3:23 description, a technical specification,3:26 or a game design document. According to3:29 my game studio concept, I need to fill3:31 out a file called game.md. I didn't3:33 have enough imagination for more, of3:36 course. You know perfectly well how I3:38 write these prompts from my previous3:41 videos. So I decided to use another one3:44 of your tips and try a skill called "3:46Grill me." You can also find this skill3:51 on GitHub or on the site of its creator3:53, who is called Matt Pocock. I just3:57 asked Claude to install this skill and3:59 sent the link to the GitHub repository.4:02 There’s a lot of stuff there, but4:04 Claude installed exactly the skill I4:06 needed, and nothing else. Next, I open4:08 the terminal with Claude Code in Orka.4:12 Actually, it doesn't matter right now4:14 where you run this. I call the "Grill4:17 me" command to access the skill. And I4:21 write: "I need to fill out this Game.md4:24 document." Then Claude conducts4:27 research, forms a list of questions,4:29 and starts asking them. Each of these4:32 questions relates to specific parts of4:35 this document in one way or another and4:37 helps to fill it out for you. In each4:40 question, there are multiple answer4:42 options, well, at least more than one.4:44 And Claude always recommends one of4:46 these options to choose. I don't have4:50 the goal right now to dive into the4:52 details of refining this prompt, so in4:55 90-95%of cases, I choose the4:58 recommended options. There were 245:00 questions in total. And after Claude5:03 asked those 24 questions using this5:05 skill, I had a finished, filled-out5:08 Game.md document on my disk. And before5:11 starting game development in Orka, you5:13 need to enable orchestration.5:15 Orchestration is currently an5:17 experimental feature. This is5:18 explicitly stated in the settings. To5:21 enable it, you need to open the5:22 settings, find orchestration, and just5:25 turn it on by clicking the "Install"5:27 button. Next, the program will analyze5:30 your environment and suggest a specific5:33 command to run in the terminal to set5:35 up the orchestration skill for all the5:37 agents on your computer. After that, I5:40 just agree with the automatically5:42 selected options. It asks again which5:45 specific agents you want to install5:47 this skill for, how to link this skill5:49 to the project, and so on. And then we5:51 just press the "Check" button. And it5:54 should show that the orchestration5:56 skill is successfully installed and up5:58 to date. That's it, after that6:00 everything is ready, and we can start6:03 developing our game using Game Studio.6:05 If you like what I do, you can support6:08 this channel on Boosty. You can see the6:10 QR codes on the screen; you can6:12 subscribe or make a one-time donation.6:14 The links will also be in the6:15 description and in the comments under6:17 the video. And we continue. To start6:20 the game development, we just need to6:22 open the code in Orca and run the6:25 command called startprompt.m. Well,6:29 basically, to execute the prompt from6:31 the file. This prompt contains all the6:34 instructions needed for the6:35 orchestrator to manage the agents. It6:38 also includes links to all the6:40 necessary documents for game6:42 development and a direct link to the6:44 Game MD file, which contains the6:45 concept of the game itself. And the6:48 first thing the orchestrator does is6:50 read all the documents and launch a6:52 couple of agents with the architect6:54 role, working on the Fable 5 model, to6:56 work out the basic game architecture.6:59 It also launches a game designer to7:02 work on mechanics, and then immediately7:05 a GPT 56 Sl, specifically a UI designer7:08, to create the basic design concept of7:11 the game. Looking ahead, I will say7:14 that it will make the first version as7:16 a prototype, meaning completely without7:18 graphics. But nevertheless, it starts7:20 creating some concepts in the very7:22 first iterations. You can take a look7:24 at them. And during the game7:26 development process, while our agents7:28 are running, we can communicate in7:30 parallel with this producer and throw7:33 some ideas at them. I mean, very often7:35 I would come to it and say: "So, the7:37 designer made some nonsense, I don't7:39 like it at all, let's redo it." And it7:41 would look at what was ready, what was7:43 done in the previous iteration. Then it7:45 would launch a new designer session and7:47 give it a prompt to redo it. Something7:50 like the studio owner said they didn't7:52 like this and that. Let's redo7:55 everything. Actually, it's quite funny7:57 because there are moments when you are7:59 away from the computer for a long time.8:01 Then you come back, and there are 108:03 different neural networks working on8:04 different tasks. You can ask the8:07 producer, the orchestrator, what they8:09 are even doing right now, and get a8:11 snapshot of the game development: what8:13 has already been done, what is in8:15 progress, and what will be the next8:17 step. And by the way, one day such a8:20 case happened to me. I left the agent8:23 on overnight on the very first night,8:25 came back in the morning, and asked the8:27 producer: "Well, what is our status,8:29 what has been done, what are you doing,8:31 and what are you going to do next?" To8:34 which the producer replied: "Damn, our8:36 development is going so great right now8:38.""70%of everything we have done is8:41 documentation and tests." And I then8:43 said: "Well, guys, of course, you're8:45 doing a great job, what can I say, but8:46 let's probably make a game." And after8:49 that, I spent several hours adjusting8:51 the producer and role prompts so they8:53 would focus primarily on actually8:55 making the game. They were writing8:57 tests, various wrappers, and useless8:59 documentation that becomes outdated9:01 with every iteration. By the way, a fun9:04 fact. I remembered myself at one of my9:07 first jobs when the director came into9:09 our office and asked: "So, how is the9:11 product doing?" And we, the programmers9:13, said in unison: "Everything is great,9:15 we've written so many tests." So now9:18 everything is just really cool. You9:20 should have seen the director's face9:21 back then; he was outraged for a long9:23 time about why we were writing tests9:24 instead of making the product. Oh well.9:27 Let's get back to our game development9:29 studio. After I redid all the prompts9:33 and restarted development, I also9:35 discovered that after one night, about9:39 45%of my weekly Claude limits on the $9:42 100 Max 5X subscription had been eaten9:45 up. And all this happened in literally9:48 24 hours. Yet the game wasn't ready in9:50 any form at all. I would run it, and9:52 there were just some red and blue9:54 squares that couldn't really do9:56 anything yet. And then I thought that9:59 for the sake of this project and such10:02 cool content—since I was actually10:04 planning at that moment to make a full,10:07 finished game that people could play—10:10I needed to shell out for the Claude10:12 Max 20X subscription, which is the $10:15 200 tier. Actually, I had very high10:18 expectations for this subscription. I10:20 expected my limits would stop running10:22 out altogether. I mean, at the very10:24 least, according to the description,10:26 there should be four times more than on10:27 the Max 5X subscription. But in reality10:30, that's not the case at all. There are10:32 indeed four times more five-hour window10:34 limits. And my five-hour window now10:38 ends with about ten, fifteen, maybe10:41 twenty percent usage, while the weekly10:44 window continues to fly by very quickly10:47, heaven knows where. I started10:51 searching on Reddit to see what people10:53 who switched from the 5X to the 20X10:55 subscription were saying. And yes, I10:58 wasn't the only one having this problem11:00. Indeed, only the five-hour window11:03 limits are four times higher, while the11:05 weekly window is only about one and a11:07 half, maybe at most two times larger.11:10 And yes, you heard that right, it11:12 turned out that in about 3 days I used11:14 up 85%of the entire weekly window on11:18 the 20X subscription for $ 200. But11:22 overall, I'll say that it does provide11:24 some freedom in using Claude. And11:27 nevertheless, the limits were enough11:28 for me to make the game. Yes, perhaps11:30 with the involvement of other agents as11:32 well. Actually, at some point GPT 5.6 S11:36 started coding more, but I wouldn’t11:39 trust it with anything other than11:41 building game interfaces and some11:44 mechanics that were somehow related to11:47 visuals or the UI. But at the same time11:52, the code was always reviewed by OPUS11:54 5, meaning after every coding task,11:57 OPUS 5 would review what GPT 5.6 had11:59 written. When I was still running out12:04 of limits very quickly, I also launched12:06 a few agents on Sonnet. Honestly, I12:10 don’t know how much better they12:13 performed compared to OPUS 5 or GPT 5.612:15, but they definitely solved some tasks12:18. They did it quite fast, and overall,12:21 there were actually a lot of bugs12:23 related to the visuals in one way or12:26 another. I suspect it was specifically12:28 with the part that GPT was working on.12:30 Interestingly, I created a tester role12:34 using GPT 5.6 Luna High. And recently,12:37 OpenAI lowered the price for this model12:40 quite significantly. So I was able to12:43 run up to 10 game testing sessions12:45 simultaneously. I literally asked the12:49 orchestrator to launch 10 playtests at12:52 once for various parts of the game. It12:56 would do it, and all 10 agents would12:58 test different sections of the game in13:00 parallel. They either did it in the13:03 browser—you could open a specific tab13:05 with an agent, see the open browser tab13:08, switch to it, and watch the agent13:11 literally clicking the mouse and doing13:14 things in the game. Sometimes they did13:17 it via scripts; they wrote scripts like13:20 Playwright to check certain mechanics13:22 using code and then provided a report.13:26 Testing methodologies were also13:28 documented somewhere so the13:29 orchestrator wouldn't have to tell the13:32 agents the same thing over and over. So13:35, certain testing harnesses were13:37 already set up there. And when we13:39 launched an agent for a playtest, it13:41 already knew exactly what to do and13:42 where. At some point, I asked the13:45 orchestrator to create a folder for13:48 tasks, which I called Tasks, and asked13:51 all agents working on the game to13:53 create MD files there with descriptions13:56 of tasks that needed to be done. And13:59 the orchestrator would then check these14:01 files, prioritize them, and send them14:03 into development. So, a lot of tasks14:06 were created for technical debt, some14:09 tasks with ideas, some with concepts14:12 from the game designer, and a whole lot14:14 of bugs that testers filed. Literally,14:18 the result of each tester's work was a14:21 bunch of bugs filed in this tasks14:23 folder, and then we would distribute14:26 them to agents with the orchestrator in14:29 batches of 5–10 items. Both OPUS 5,14:33 GPT 5.6, and Sonnet 5 fixed these bugs.14:37 By the way, on the third day, once I14:39 had more or less set up this pipeline14:41 for my game studio, I asked the14:43 orchestrator: "Why are we taking so14:45 long to make the game?" We started14:49 looking into it and realized that doing14:52 one task at a time really stretches14:55 things out, because a task needs to be14:57 picked up, researched, implemented,15:00 reviewed, and then tested by a tester,15:03 which takes half an hour, maybe an hour15:06. So the orchestrator and I came up15:10 with an approach where we accumulate15:12 tasks for a while and then distribute15:14 them to agents in batches of 10 or 20,15:17 depending on the size of the tasks.15:20 Essentially, a mandatory criterion for15:23 describing a task became its size: S, M15:27, L, XL, and so on. And a rule was15:30 written on how these tasks are assigned15:32 for development. That is, no more than15:35 one XL task, no more than three L tasks15:37, and so on. By the way, I have a15:39 Telegram channel; you can see the QR15:41 code on the screen. There I post15:44 updates about new video releases and15:46 sometimes additional reviews or links15:48 to games—things you won't find on15:50 YouTube. Subscribe so you don't miss15:53 out. Now, let's continue. By the way,15:55 here is another interesting fact. At15:58 the very beginning, I mentioned that15:59 all agents launched in this16:00 orchestrator have access to the system16:02 functions of this orchestration16:04 environment. Literally, every agent16:07 sees the current state, how many16:10 projects are running, how many working16:13 trees are open, which tasks are ready16:16 or not, and even things like the16:19 remaining five-hour window limits or16:22 weekly limits. It was actually16:25 monitoring how limits were being spent16:28 on various tasks and could make16:30 decisions based on that, whether to16:33 launch more or fewer agents, or assign16:35 them more complex or smaller tasks.16:38 Once again, the link to the GitHub with16:41 my game studio will be in the16:43 description below the video. Go grab it16:46 and feel free to try it out. Let me16:48 know in the comments if you liked it or16:50 not. So, in the end, after 4 days and16:54 90 hours of continuous agent work, the16:56 game was finished. And before we look17:00 at the result, here is some statistics17:01 and numbers, if you're interested in17:03 that sort of thing. I asked Claude to17:06 gather statistics on the current17:08 project with interesting facts and17:10 create this landing page. The game is17:12 called "Hero with a Sword in a World of17:14 Magic." It is a turn-based strategy17:17 game about a hero, a city, and an army,17:19 assembled by an agentic studio in 417:21 days and 10 hours. "Barely slept," it17:24 says. We made this game from August 8th17:27 to August 12th. 102,700 lines of code17:30 written by Claude Opus. And here is the17:33 image from the title screen. 9 hours of17:36 silence over four days, which is17:39 literally 90, well it says 91%of the17:41 time the agents were working. 106 hours17:46 of activity on the project were logged17:50 in total. Only 9 hours of downtime. 617:53 hours of which were on the first17:54 evening. I don't know why. Maybe I ran17:56 out of limits there. On average four17:59 agents worked simultaneously, peaking18:01 at 30. That was most likely during the18:04 playtest sessions. As I mentioned, I18:07 was running 10 testers simultaneously,18:10 while developers fixed bugs in parallel18:13. You can see the chart here. The18:15 spikes for these agents happen at night18:17. That is, between the 9th and 10th,18:20 and between the 11th and 12th. The18:22 final night, which was the crunch18:24 before the game's release. A total of18:26 543 agent sessions were launched. One18:29 of them lasted 70 hours—that's the18:31 orchestrator. You can see the root of18:33 the project here. That's the 70-hour18:36 orchestrator session. And there are two18:38 more sessions in the project root. This18:40 is the final polish. I polished the18:43 game twice. After all the agents18:46 finished working, we used Fable 5 to18:49 finalize the result. You can see here18:53 there were 311 Claude sessions, 232 in18:56 Codex with a median of 17 minutes. And19:00 a maximum of 70 hours for the19:01 orchestrator itself. 13.6 billion19:04 tokens were spent on the project. There19:08 is a breakdown by model here. You can19:10 see it. 98%of tokens are cache reads,19:15 1.5%are cache writes. And you can see19:19 that 0.5%is the entire novelty of the19:23 project. That is, literally, all the19:25 code that the agents wrote. The cache19:27 saved $ 57,000. This is the most19:30 interesting part. Converted to API19:32 costs—and I remind you that I use19:35 subscriptions, but if I had used the19:37 API, I would have spent $ 7,500 on OPUS19:40. On Luna, I spent only $ 21. And I19:45 remind you that several times I ran 1019:48 playtests simultaneously. And a great19:51 many tasks, especially milestone ones,19:54 were also tested by Luna. And GPT-4o. I19:58 have no idea where it even came from,20:00 but one session went through there.20:02 This is how much of the subscription20:03 windows were consumed. 85%of the weekly20:06 window was eaten by Claude. 52%by Codex20:09, and we have a separate usage window20:12 for Fable 5, which was 21%used. Luna20:16 did more than half the work of Codex20:19 and cost 30 times less. That's how much20:23 cheaper Luna is than GPT-4o. In reality20:26, $ 0 was paid for the API. As I said,20:30 I worked on subscriptions. How they20:33 wrote code at the peak: a commit every20:35 99 seconds around the clock. That means20:38 for an entire day, the agents were20:41 committing to the game repository every20:43 99 seconds. A total of 2,360 commits,20:47 about 102,709 lines of code in the game20:50 itself, and the rest went into20:53 documentation. 499 tasks were completed20:58. You are probably already curious to21:00 see what kind of game turned out. I am21:02 actually quite curious myself. Five21:04 checks for every single edit. This is21:07 how the agents double-check themselves.21:11 From bare HTML to a spellbook in 3 days21:14. And there is horizontal scrolling21:16 here. You can see the game's21:18 progression here, how the graphics21:19 appeared. All graphics were made by GPT21:22 5.6 S using the Image Gen built into21:24 the codebase. That is, it generated21:27 these sprite sheets. Some of them had21:29 animations, some didn't, and some21:31 animations are procedural within the21:33 game itself. And bit by bit, a game21:35 like this came together. There is even21:38 a timeline here showing when certain21:40 parts of the game were made. Like, it21:42 was working on the game's style on the21:45 night of August 11th. And this is the21:48 game map that appeared. It says from21:50 placeholders to landscape. There is21:52 already some basic graphics here, as21:54 you can see. So this is not the21:55 earliest prototype. And here is a21:57 screenshot from the final game. Though22:01 it looks a bit distorted. How did it22:02 all end? Acceptance shots from August22:05 12th. And now I suggest we play the22:07 game and see what we've ended up with.22:10 Launching the game. And here is the22:12 start menu. The game is called Hero22:14 with a Sword in a World of Magic. You22:16 can load a game here, there is autosave22:18, you can switch the language from22:21 Russian to English, as well as toggle22:23 music and sound effects, which, by the22:25 way, were made entirely procedurally. I22:29 am starting a new game, and each22:31 session starts with a random plot.22:34 Overall, at the very beginning of the22:35 game, we have a choice. And I22:37 understand that this choice somehow22:39 affects the composition of our hero's22:41 starting army. Then we immediately22:45 choose our hero's first-level skill and22:47 end up on the map where we can move22:49 around. The game very much reminds me,22:52 well, perhaps of Heroes II visually.22:56 First thing, I will move into the city.22:59 And in the city, you can hire more23:01 troops and build various structures.23:03 Then, resources are scattered across23:04 the map. These resources can be23:05 collected. When picking up a chest, we23:07 can choose what to take: gold or23:09 experience. The first battle. And this23:11 is what the battle screen looks like.23:13 You can move around and attack23:15 opponents. There are ranged warriors23:18 and melee ones. Also, there are23:20 obstacles scattered on the battlefield,23:22 which add variety to the combat itself.23:25 After the battle, we gain experience.23:27 Along with experience comes a new level23:29, and you can choose new skills. When23:31 the hero's movement points are spent,23:33 you can end your turn. Next is the23:35 enemy hero's turn, who, by the way, is23:38 in the same situation as we are. He23:41 cannot see us. The fog of war affects23:43 him too. He needs to explore the23:45 environment just like we do. Search for23:49 resources, find mines, capture23:51 objectives, and look for our location23:54 to defeat our hero and take the city. I23:57 am engaging in a series of battles.23:59 Overall, it is quite easy to play. Our24:01 level is already quite high, and I24:03 hired plenty of warriors at the start.24:07 So, in principle, dealing with the24:09 neutral units on the map is not24:12 difficult. You can build in cities24:16 without visiting them by24:18 double-clicking, as well as recruit24:20 troops directly so the city doesn't24:23 stay empty. Otherwise, if an enemy24:26 appears from the fog of war, they will24:28 simply take the city, and we will lose.24:31 By the way, I made a demo version of24:33 this game. You can play up to the tenth24:36 turn. I have uploaded the demo version24:38 to. The link will be in the description24:41 and in the comments below the video.24:43 Meanwhile, we are hiring some neutral24:45 creatures on the map. These are24:47 horsemen running backwards, and we24:50 fight neutral units for resources.24:53 Overall, the game loop is structured24:54 something like this. We need to explore24:57 the map, we need to gather resources.24:59 There are different types of resources.25:01 There are crystals, wood, gold, and25:04 something else. These resources are25:06 needed to build and upgrade structures.25:09 When we build and upgrade structures,25:13 we can recruit new, stronger creatures25:16 into our army to better explore the map25:19 and defeat the enemy on the other side.25:23 We are slowly moving through this game25:26 loop. By the way, the game also has a25:28 quest book. Contracts periodically25:30 appear there that need to be completed.25:33 Completing these contracts gives gold25:35 or experience, and sometimes units for25:37 the garrison. There are main quests25:39 that must be completed according to the25:41 plot. And the game even has some slight25:43 narrative variability. As I said, there25:46 is random dialogue at the start, random25:50 dialogue at the end, and along the way,25:53 events occur that influence the story25:56 in one way or another. I believe there25:58 were about 10 branches and two endings26:01 implemented in total. Meanwhile, we are26:04 steadily advancing toward the enemy.26:07 And here I have found his castle, which26:09 we are about to capture. The enemy, by26:11 the way, has a decent squad, quite26:13 comparable to ours. During the battle,26:15 a combat log is written, recording26:18 events, who hit whom, and who moved26:20 where. There is, by the way, a26:22 spellbook. You can learn and use26:25 various spells there, but in this game,26:28 I focused entirely on our hero's26:30 defense and attack. And so, we won. We26:34 also lost quite a few creatures, by the26:36 way. And this is what the victory26:38 screen looks like. On the eighteenth26:40 day of the third week, we captured the26:42 enemy city. The sword struck before26:44 winter arrived. You took their city26:46 before the roads became impassable. By26:48 the way, it looks like ChatGPT wrote26:50 these texts. Let me know in the26:52 comments if you liked it or not. The26:53 link will be in the description and in26:55 the comments under the video. You can26:56 try playing this game yourself.26:58 That’s the video we have for today.27:00 And this is our experiment with agent27:03 orchestration. Hit the like button if27:05 you enjoy videos like this, and don't27:07 forget to subscribe to the channel. It27:09 helps it grow. And goodbye, everyone.27:12 Uh.
4,360 words · 646 lines
