Fable 5 & Qwen 27B – Traycer Multi-Agent Hands-On Test!

We're starting with Fable 5 and I've told that you're going to be directing Quen 3.627B to build a realistic flight simulator game. Today we're going to be looking at a really cool opensource multi- aent orchestration tool from Tracer AI. So this we're going to be reading at least the introduction for this just off of GitHub right here because this is open source and MIT license. But this has some really interesting features. And if you are interested, as has been becoming more and more popular in having bigger, more expensive models kind of delegate tasks to smaller, cheaper models, kind of like sub agents, this is something that does it very very smooth. Smoothless. Smoothly. I was trying to say seamlessly and smooth, and I said smoothless. Nonetheless, for a smoothless experience, I currently have this working right now. Just going through a code repository that's on the system for an Arduino ray tracing rendering cluster and we can see right here that GPT56 soul on low is delegating tasks both to DeepSeek V4 flash just running here within Open Code and it also delegated out a task to sonnet 5 running from within cloud code to basically make a website for this repository. There's a bunch of cool features here and we're going to be taking a look at this today. And I think it may be fun additionally to see how a local model acting as a sub agent performs when delegated by a couple of big state-of-the-art models like Kimmy K3, Fable 5, or GPT56 Soul. And then we'll use a trusty Quen 27B to actually show some differences in performance based off of that. But all using the Tracer desktop app right here. So before we get into it, I want to say thanks to Tracer for supporting the channel. I have worked with them in the past and they allowed me to do a video using one of their other systems which was basically like spec driven development and we ended up making something cool. So that's still on the channel if anyone wants to look at it. But additionally to that, if you want to subscribe, please feel free to do so. I'm trying to get that 100K plaque. It would look good on the wall behind me. So let's take a look at some of the interesting features here. One that I personally like very much is the first one listed as bring your own agent or BYOA. Basically what we saw when the app was open at least taking a preliminary look in the introductory section every agent that is open right here GPT56 soul and sonnet 5 these are just using my own systems native codecs install and the native claude code so all of the authorization authentication everything is just happening through my personal subscription with either of those openai and anthropic and then additionally as we saw right down here well this is just using open code go I believe for deepseek for free high, but we're going to later on use a local model just running on the system behind me. And that's pretty cool. So, it's pretty easy. I mean, you don't have to like you don't have to, oh, I have to pay Tracer and then they sell me the tokens to these models. I can just natively use my own agents for this. And that's what they mean by that bring your own agent. And that makes things a lot simpler. Now, the next thing in this list of features I want to touch upon that is pretty cool is the agentto agent communication where it can create the automated loops where agents can speak to one another. But if we click on this linked compatibility matrix right here, it opens up a documentation page with a bit more information on this agentto agent feature. And it's really kind of cool because basically the agents will coordinate with one another all from within this one single window in this task. And beyond that, basically as we see right here, just take mental note of this. This highle thing right here that says repository summary request was created just with GPT56 soul and everything under that. These are all sub agents that it spawned using different coding tools. So when they say bring your own agent, some of these were spawned just in clot code. Some of them were spawned in open code. So, it's pretty cool how this one agent right here was able to spawn and delegate tasks to all of those just by me speaking to this one agent. And when we go back in the docs, we basically see under what it does is a lot of what we just saw there. All of these capabilities are able to be triggered by any agent that we're using. Assuming it's intelligent. I mean, if you have like a local agent hooked up to like a old llama model, you may encounter some issues. But assuming decent intelligence like that 56 soul on low or sonnet, it can do all of this. And this is just another visual representation of what we saw right there with the spawned sub aents from different agentic coding tools were actually created underneath the big one. They also do have cross device sync. So if you want to swap from between systems, it is useful to have that capability. That's something that's become big with a lot of coding tools lately is just being able to access them from more than the singular machine they are running on. Additionally, speaking of machines and being run on, this seems to have very good software support because it's not just currently Mac only. There is an Ubuntu install, even different flavors of Linux, which is nice to see. Additionally, Windows is there as well. Now, for this video and today's demonstration, this app is running natively on my Ubuntu system, and it's always nice to not have to swap over to a Mac to kind of install and play with newer and cool things. So, I very much applaud that wide compatibility for different OSS. Now, as we can see something that I had done and my introductory example right here is just starting with this looking through an Arduino render form thing that I had made with Fable 5. And we can see that GPT56 soul here had spawned multiple sub aents just from within open code to get a performance analysis of the actual software and the Arduino being linked together and running. And we can see when agent A or analysis A had completed which was running from DeepSec V4 flash right here. it just gives us back its main conclusion. Analysis B is still independent and will return separately. So if we actually and the UI here is really neat. So I can just drag that as basically like a second tab there. So it's almost familiar because using a web browser has a lot of similarity here. And this one was using Kimmy K3. So it's going to be pretty interesting to see just the way that well obviously this will take significantly longer being that that model likes to think a lot. This is actually a non-trivial level of complexity. this little Arduino ray tracing repo that it's going through right here. But it's just pretty cool to see and like you can manage a lot of different agents and tabs here from within this one interface. Now, additionally to that, we also do have these artifacts right here which are shared. So agents can basically leave information if we want them to added on to these artifacts. Therefore, if we're going to just need like a TLDDR or scratchpad or something like that, these make it very useful. And again, in the same way, we can essentially have these open in different tabs, move them over to different windows. So, if I wanted all my artifacts somewhere in one specific pane, we can see we have them right here. Now, in terms of more like UI tour and things of the sort, something really cool to see here is this almost performance monitor looking thing. This actually shows us our usage limits across some of the different coding agents that we have right here. Now we can see that fortunately for my GPT codeex weekly session I'm at 0% used and that is actually correct. They have been handing out resets very favorably lately to those subscribers. But if we look at cloud code we can see my current session has used 4%. And I do have some other tasks running on different machines using some cloud models. So if that seems to creep up a little longer than what we would expect based off of what we do in this video that is attributable to that. Additionally, we also have a little CPU logo right here. And if we click on that, it just shows us the system resources that this is using. This is nice to see, especially because we're using an Ubuntu system. And it's not often always the case that you get tools like this that work seamlessly and are nice. So, I like to be able to see that as well. And it just shows us like CPU, memory, RAM share, etc. Other things just for a little more familiarity with the UI, we have open terminals if we do have any that we can select and go through from within there. We also have git diffs if we do have a specific repository that we're working from in. We have a file tree so we can see all of these specific files and assets from within our current working project. And then finally, we have the ability to share this. So I could basically enter in someone's GitHub handle or email and it would be able to share this project project with them. A lot of stuff that's mentioned here in the GitHub is basically just talking about the ability to share this across teams. out for collaborations if you are working in a team and not just a solo founder as myself I think is the term also then we can just kind of close this out and have a little more screen real estate just to monitor our agents and Kimmy is still working on its performance analysis for this rate tracing render. I'm going to let that finish out and then we'll probably have it give us some summary or Arduino artifact here and then we'll move on to probably seeing how Quen 3.7 3.6 627B performs when being delegated as a sub agent from a bigger state-of-the-art model. And we received our view of the second analysis, which was run by Kimmy K3 versus the DeepSeek V4 Flash model. And we can see our main agent that's coordinating all of this right now told us the second audit is the stronger quantitative analysis, which definitely tracks based on model size and performance, etc. But it's also interesting here just to be able to almost compare the results of different models just from within one main interface right here as it wasn't even my original intention. But additionally, what we just got is a comparison of the same task from within DeepSeek V4 Flash and then Kimmy K3 where they need to go through a codebase and give us a performance analysis of some more intricate and lower level code. The last thing I'm going to do just in this first test, which is more designed to get a feel for the interface and get warmed up and things like that, is tell 56 Soul to have Sonnet add the performance analysis, the good takeaways that we received to the website for this specific repository. So, we're going to see in the pane right here to the right, Sonnet 5 is running, and this will just delegate the task over to Sonnet. Following that, I've also said that I would like some of these artifacts updated just with where we specifically are at the current state. And we can see here that I actually made a bit of a mishap. So I had neglected to realize that it would have opened sonnet right here in a new sub aent task and then told it to do that. So now with that open, we can see that it is currently just updating the website with the performance claims that we do want to keep. And now that Sonnet has finished the task that was delegated just in improving the website with what we learned from the Kimmy and Deepseek assessment, it's going to update two of the Tracer artifacts to reflect what it's done. So, if we open back up the side panel right here, we'll see that the artifacts are contained just down here in the artifact section. And I do have most of them open right here. So, we'll allow it to do it. And it also does just give us a hyperlink of what it did. And if we open that, we see it opens up here in a new tab. So, I'll just drag that down and we'll be able to see it right there. And it also does exist as we see right here. And part of this is what it had done before when it originally made the website. And then part of it now as we see the performance analysis reconciliation. Let me move this up cuz my head is sometimes down here in the video. So we can see that the performance analysis reconciliation was added in following this after the assessments from Deepseek and Kimmy. And we can see that it does receive the message that the agent has been completed here in the main one. when we begun it from. So basically GPT56 sold saw the message that Sonnet 5 has finished and it received that right here. So basically it's pretty much complete. It's just going to wrap up whatever it needs to in terms of its job here kind of acting as our orchestrator. Then I suppose we'll take a quick peek at the website because why not? And it was a cool project that this is actually based on rooted in and working on. Then we'll get into perhaps some more entertaining multi- aent conversational or sort of things. So, here's the updated site for the Arduino render cluster, and this was made with Sonnet 5 on high thinking mode, just based off of some of the information it gathered from our other agents. And when we scroll down to see the performance analysis section, we're going to see a lot of what Kimmy K3, probably more of that than what DeepS V4 Flash gave us. But, nonetheless, it's interesting to see that this product right here or website is the culmination of multiple different Asians working together. 56 Soul, Sonic 5, Kimmy K3, and then DeepSeek V4 Flash. So, it's pretty interesting. It even has a nice wiring diagram and things like this. And these are all true to the specific board layout. And this Arduino thing was kind of maybe not as visually exciting to see, but kind of cool just in terms of like technical. A final thing I want to touch upon before we get into some more exploratory playing with this is the different modes that are available. So, while we were performing all of these tasks, you may have noticed that the main agent right here, which was 56 soul, was shown with this epic mode enabled, while all of the other agents were just shown in regular mode. So, if we go back to the docs right here, we can see that regular mode is direct work with an agent. So, basically when Sonnet was told to make the site better or when Kimmy or Deepseek were told to do a performance analysis of the repository and then epic mode is Tracer structured planning layer around the coding agent. So, that makes more sense now. why that was enabled for the GPT 5.6 soul agent which was kind of orchestrating all the other ones for us. So next up we're going to be starting just here with a blank slate and I can change between whether I'm using sonnet or any of these other coding agents from right here when we initiate our task. I think for this I will probably actually use Fable just because I'm interested in seeing how well it can do when paired with a locally running Quen 3.627b. So I think that'll be pretty entertaining and I'll use it on probably high should be sufficient for the task. Additionally to that I am also going to select a directory that we're going to be working from within. I did just make a new empty one. So whatever we do end up doing with these model combinations it will have its own little place to live. So they now have a folder added. I just named it multi-a for multi- aent. I know that is very very creative. And we also do have epic mode enabled because if we recall it just makes more sense when this is going to be used primarily as an orchestration agent and things like that. So I want to do just some AB testing to see how well the same exact model can build something when given a different orchestration agent. So we're starting with Fable 5 and I've told that you're going to be directing Quen 3.627B to build a realistic flight simulator game. The reason I want to use this prompt is because visually we'll be able to tell the differences between how this orchestrates that model versus how 56 Soul does or something else. And it's also just kind of fun to play. It's going to begin from the pilot's POV in a passenger jet starting on the runway. It should have interactive flight controls, detailed scenery of the ground, and be overall well done. You are not going to help code it. You are going to guide it as a skilled adviser would do. And then I'm telling it the quen model is being used through open code because I have that one selected right there. And we can see it did just actually create this sub agent for open code which has that model already selected. So I'm happy to see that. We can also drag these in different. So we see like right now it's a much bigger workspace. So I could just put this one on the bottom right there if we wanted to. And that probably works well for the purpose of this demonstration. And I will close that side pane for now. So, we can see that this received the message. You're the builder for a browserbased realistic flight sim. I'm your technical adviser. I'll give you the architecture in one milestone at a time. Review your work and then we can see it basically enhanced the prompt and is going to make this a bit more interesting. I'm wondering should I do these split screen or top to bottom? Probably side by side because then we get more vertical real estate instead of less. So, I should be able to just do that. Good. And now we have these side by side. So Fable 5 on high on the left here orchestrating our local Quen 3.627B which is running on the PC behind me. So this may take a little bit of time. We can see that this is going to wait to monitor process for the builder. We can also see a bit more information about how Fable 5 is deciding to approach this saying rather than handing it the whole game at once which it would likely thrash on. I locked it down up front and I'm giving it one milestone at a time with acceptance criteria reviewing between each. So the current milestone which is a sign now that this is working on is a static world sky sunlight 20 km ground plane marked in 3,000 meter runway cockpit height camera at the threshold render loop and a placeholder instrument bar showing speed altimeter heading and throttle. All right and then we have additional milestones assuming that it does a good job on that. And we also get a bit more information just about the specific model. So we can see right here this is the BF-16 version of this running locally. So, I'm using a Blackwell 6000 Pro cart to serve this. All right. And we can see right here that our sub agent has completed milestone 1 and it sent it back now to Fable 5, which is going to review it. The builder reports milestone one is done. Before assigning the second one, I'm going to verify the claims by reading the code. It can't open a browser itself, so zero console errors needs checking. Interesting. And this is partially why as models become more intelligent but also more expensive to run at the highest levels of intelligence, being able to have them used less frequently to control a cheaper, smaller, but still capable model like this is going to become more and more popular. It already has begun. All right, we see milestone one review and approved. I read every file and syntax checks them. Clean structure, correct API usage. Camera geometry is right. Nice work. Oh, okay. And it's telling us now. So that right there is the message we saw that it had sent to our sub agent right here which it has also inevitably told it to start milestone 2 which is what it's doing here. Now we can read its summary that it's just giving us the report was honest. All five files match what was claimed and it passes every module. I caught three latent problems it missed. Controls only update on key events. Depth buffer precision and key mapping conflict. Milestone 2 is now in progress. flight physics and takeoff. Since a 27B model would likely thrash trying to design aerodynamics itself, I prescribe the exact model. A point mass velocity follows the no scheme with realish constants. 60 ton jet 240 kontons of thrust. I may be wrong about that. Quadratic drag tuned for 250 m/s and speed stop speed 140 knots rotation speed corrected turns from bank angles. Okay, so basically milestone 2, it gave it a bunch of more or less proper physics things for the specific jet in this specific environment because it said the tiny model would probably not have the depth to do that. So it can verify behavior without eyeballs to the screen. Okay, cool. Remaining milestones after this, live instruments, cockpit framing, then detailed scenery, and then polish. Cool. So after quite a bit of time, we're now on to the final assigned milestone for our Quen sub agent to perform, which is the sound and feel. So this has really been going on for an incredibly long time, just with these going back and forth with this orchestrating with this performing fixes based on feedback. So something interesting Fable picked up on is the builder had correctly implemented a pitch direction despite my instructions containing an error. It followed the real instrument principle over my parenthetical, which is exactly the judgment I wanted to keep exercising. That was pretty interesting that Quen 27B found an error in the Fable hypothetical implementation and it was happy about that. Fable was interesting. So, we see right here that the agent hit a context length limit. I currently have this set to a context length of 131072. That is the Quen model running locally on the Blackwell 6000. That is because it's the BF16 version. So a Q8, yes, we could push to max context, but I find it interesting because now Fable has to decide how to work around this. Very cool. And what it's doing is basically saying, okay, the 5.1 fixes and some of the additional things are not really insanely big asks. So it's going to create a new agent. As we see right here, it has just done. So let's put that we'll just put it to the side right here. And now this brand new agent is getting started based off of what Fable's doing. So, it's interesting as well that even when hitting a context limit like that, we basically just had our model fix that the orchestrator at least just by giving a new agent some of the things it needs because one, our context is still with the main orchestrator and two, as it said there, it's not really enough like problematic stuff that it needs to do that it couldn't be knocked out by a fresh agent with a fresh context window. So, it's interesting. All right. So, unfortunately, and I want to make it clear, this is an issue with the local LLM running on the PC behind me. It keeps getting stuck here. So, at this point, I'm pretty much ready to tell Fable to just present to us the game as it is without like it won't be the 100% polish, but it'll be like 95% there. And then I'll have to troubleshoot what exactly is going on with this Quen instance because um it keeps freaking out with this one specific fix. Interesting. So, Fable has decided on its own that it is going to apply the fix by itself and then present it to us. Oh, okay. Nope. It's sending it to the builder one more time saying you've been cut off midterm twice. Both times right at the first file edit. Smallest possible steps this time. Do these as four separate small tool calls verifying after each. We'll see what happens. So, something's wrong with my local system. I'm going to have to troubleshoot that and figure out why. But the cool thing about this orchestration here is that it actually showcases almost like a fail safe where if our remote agent is not working very well or any one of our sub agents having the big smarter orchestrator can basically just get us to the finish line. So I'm saying apply the remaining fix and read me yourself directly. The builder is clearly wedged as that was a suggested follow-up from Claude. So finally before we look at it, I'm asking the big orchestrator to create an artifact on how to run the game and then also how the small sub aent model did throughout the entirety of the build. So we'll get that as an artifact and be able to look at that. And really like where this started freaking out and the servers hung on my system, so it's something I need to troubleshoot myself. It's a networking issue or something's not 100% right on the local AI box. Basically, Fable I think fixed like one or two lines of code. So everything here that was written was entirely Quen 327B at BF-16 under the direction of the Fable 5 model on high thinking. So that's kind of cool. And now we have our information right here which we can open. All right. So we have controls debug spawn test how to play it how Quen 3.627B performed as the builder. So we have the scorecard as well. Very interesting. good, notable, judgment under bad instructions. And that's something we touched upon when Fable had realized, oh, I gave it a mistake and it actually fixed it. It did the right thing instead of just following me blindly. So, that was pretty cool. And it's saying invisible in practice feature was caught by external verification, never by the model itself. Roughly 30 substantive defects were found across four fixed rounds. None survived to the final build. All right, let's take a look at it. All right, so now we have our completed result being served here. Okay, very interesting. So this is again like are there different camera views? Okay, interesting. Look how this the way the speed is actually functioning in a realistic way and the acceleration definitely shows us some of the big models implementation here because that would not be the case if this was happening with the small model. It would have just made the plane much quicker. I can tell you that as a guarantee. Okay, here's what I'm going to do. I'm going to restart this. I don't know that my speaker is set to be the output audio device. Okay. It's interesting because the cockpit model is not as good as I would have hoped it to be. And again, the entire thing here was made by Quen, but it did do the instrumentation. And we do have a city in view. We also have a runway, which is decent. Okay, we should be around takeoff speed now for this specific type of plane. How do we arrow up and arrow down? is nose. Okay, so we need to take off now. Oh, arrow down brings us up. Okay, it's inverted. That actually makes sense. All right, and we have our roll and then we have A and D for the like the tail fin thing. Make of that what you will. All right, let's just do a pass over the city. All right, it's pretty basic in terms of like the uh the city model, but it's pretty populated. Here's what I'd like to do. Let me Let's turn this thing around and see if we can land. I don't know if we'll be able to, but So, the airport is there. Oh, yeah. I don't think we're going to be able to land. Nope. We're not going to be able to land. I'm not even going to try it. We're going way too fast. We'll just do a pass by the airport then. All right, that's cool. And it did have like a like if we do crash, I wonder if we can stall. Oh, okay. All right. Overall, that was kind of boring. And I guess personally, I find some of the Flight Simulator products to be kind of boring. So, I suppose that shows this was well done. But the whole entire point of the thing was basically what we saw here with Fable orchestrating the local Quen model to actually build that which is kind of neat. So in the meantime, I want to try this just entirely with Quen 3.627B running locally on the machine behind me doing the entire task from start to finish. So we're going to start a new session here. We're going to put it in epic mode and give it a slightly modified version of the prompt. Being that the prompt had told the model that it needs to spawn a sub agent and then control it. Since this is going to do the entirety of the task by itself, we're going to just tell it basically to give us that deliverable. And we'll see how this goes. So, after about 32 minutes, it did produce a result for us. Now, I haven't looked at this yet. I don't know if this is going to be any good, but nonetheless, it does seem that it is running right now on port 8080. So, let's see what Quen 3.6 entirely by itself did. Oh, wow. Okay. Um, it's definitely a very interesting AB test of how this model does when given that big prompt in its entirety versus when given a smarter model to orchestrate it. Now, some things I'm going to notice are that actually like this did a better job with the terrain. At least it has more terrain. Actually, I don't know that I would say that. Um, I don't want this terrain for like a flying game. There's no room to take off here. Uh, there's also no sound. All right. Yeah. And unfortunately, I've tried some of the controls here for throttle and things like that, and we don't really get anything functional. So, this was definitely interesting to see cuz unfortunately that just didn't work at all. And then the one where we had Fable orchestrating Quen and we told it don't change any of the code. You just guide it to do things correctly. I also want to note that I know this may sound a little weird, but this one seemed like it had actually implemented some more realism in terms of the physics than I would have expected. And we did kind of see that when it was going through the planning stages and it was giving some like thrust values and other like force and physics calculations. So, just a really interesting AB test to see how this performs. And really, that's probably going to bring us into the conclusion here. I think that Tracer Desktop with this multi-agent coordination capability is really fantastic. The way you can even just move things around seamlessly, the way it has support for so many different types of agents, it makes things really simple and it handles like if you tell an agent, okay, go and spawn a sub agent in this, it doesn't. And it just makes things easier. And this was really quite enjoyable to use, especially when wanting to compare how models perform with other models. So if we did an entire video just testing how a Quen model performs with a bunch of different Asians controlling it. So like Fable 5 controlling Quen 27B or Opus controlling 27B. This would make that a really really seamless video to do and is definitely something to think about in the future. So that is probably going to wrap up today's video. If you have any questions, please feel free to leave them in the comments.