¶1The AI industry wants one big winner. GPT 5.6 Soul or Claude Fable 5. I'm not picking one. Picking is the mistake. The winning mindset in the age of AI is and not or.
¶2Combine compute. Don't select compute. Don't get me wrong, I maintain a model stack to track the best models. But when it comes to pushing the frontier and getting real engineering work done, you want teams of agents working for you, not an unchecked lone wolf. There's a simple yet powerful pattern for this [music] that's existed for years.
¶3Yesterday it was called architect editor, prompt chaining, then agent chaining. Today Devon open router and [music] others call it model fusion. Tomorrow, no matter what it's called after you watch this video, it'll [music] be another tool in your agentic engineering toolbox. I've been running this pattern for years across hundreds and thousands of agents. You probably have too without naming it.
¶4If you use Model [music] Fusion properly, you're combining the intelligence and the context windows of your models. [music] So, how do we tap into Model Fusion for our agentic engineering work? Once again, we don't vibe code. Instead, we harness engineer with the PI coding agent to build a powerful fusion [music] harness. There are three commands we'll use to orchestrate our tight coordinated two agent team.
¶5Opinion to get multiple perspectives from our agents. Fusion to combine and consolidate the best results of our agents and auto validate to use intelligent on the-fly validation to improve on the second constraint of agentic engineering review. We're going to start with a couple of powerful workhorse models cloud sonic 5, GPT 5.6 6 Terra and then we'll move up to the state-ofthe-art models. Let's start simple and move to the most valuable/ opinion lists the top three scikitlearn models, use cases, pros and cons. So when we fire this off, you're going to see exactly what you would expect from a fusion harness.
¶6We have two agents working for us in parallel. This is two unique perspectives that we wouldn't get out of the box with any other agentic coding tool. So if we scroll up here, you can see exactly what happened. And in fact, this tool is already giving us insight into how these models perform side by side. Terra blew through this problem in just 4.5 seconds.
¶79K tokens in, 300 out, 3 cents. Cloudset 5 took double the time. 4K more tokens in, 4K more tokens out, and it costs about a cent more. We're nitpicking this small prompt, but the key is that you can see your agents side by side. Relativity is the best benchmark.
¶8You can see we have a nice breakdown here of the top scikitlearn models. Random forest should be the top answer and then they both list a couple other ones. So let's go ahead and combine the results into one. One agent response is fine. Two agent responses is good.
¶9Combining the perspectives of two models is even better. Let's do that now with /fusion as table.md in your current working directory. So once again, we're going to use two models to do this work. They're both going to execute on this, write their table so that we have a concrete result. But then something different happens here.
¶10After the models do their work, a new fusion agent using the same model as our architect is going to pick up this work and combine the responses. So this is the fusion prompt applied directly to agents in a dedicated agent harness. Now we get to see where the models agreed, what's the consensus, what's the divergence and what was discarded. This information is ultra ultra valuable. Okay.
¶11So let's just understand the results. What are the best scikitlearn models according to sonnet 5 and gpt 5.6 terra both agents completely agreed here. This is great random forest grade in boost logistic regression. And you can see their explanation of why. As engineers when you're making decisions you want to understand when to use technology and you want to understand the pros and the cons.
¶12Engineering is all about trade-offs. Agents don't change that. At the end of the day, you have to collapse your work into a decision. And if you want to make the best decisions, you want the best compute available, giving you their perspective, giving you their input, giving you raw informations you can use to build valuable software. You can see we have a consensus.
¶13They both all agreed on all these. We had a small divergence. Both answers were complimentary, not contradictory. So, this is good. Our fusion model kept both and there was nothing discarded.
¶14You can imagine how powerful this can be when you're making game time critical strategic decisions for your engineering work, but also for your business. Once again, we're scaling our compute to scale our impact. Now, let's use the most powerful custom slash command we've [music] built into our fusion harness/auto validate. implement an MVP of these models with the smallest possible data set to showcase their respective pros, cons, and pros, consy, same dur. So, super simple example we're working through here just to showcase the power of this fusion harness.
¶15What's happening here? This is a unique agentic workflow and it's built into the fabric of the agent harness. When I type /a valvalidate, a workflow inside the harness kicks off that does this. You can see here the builder is just waiting. GBT 5.6 6 Terra is doing nothing.
¶16What is it waiting for? It's waiting for the validator to create a validation gate. Cloud Sonnet 5 is building out a validation gate right now that proves that the work is done. With this fusion harness with powerful workhorse and especially state-of-the-art models, we can do something really incredible. Any prompt we write can have a validation flow written before the work is even started.
¶17The validator is done now. It's written a raw validation script that proves that the work is done. It also has fail commands to give feedback to our builder. Our builder has very very quickly built this out and really impressed with the speed of the GPT 5.6 Terra and Luna models. Now what's going to happen is really really cool.
¶18So the validator immediately executed the validation script. And you can see there was no file created at all. And so the validation failed, right? It was just running this as its initial check. But now this validation script is going to run against the work from the builder.
¶19So check this out. See if we got any fails here. Pass pass. All passes, no fails. Big shout out to 5.6 Terror.
¶20Really powerful workhorse model. Kind of sitting right next to Cloud Signet 5. I think it's actually outperforming it by a bit, especially when you take into account the speed in which it gets work done and the concision in which it does it. This is all good. We can scroll back up to the validation script that Sonnet 5 put out for us.
¶21Like look at all the validations. Every one of these fail statements is feedback for our builder agent if something goes wrong. So you can see it's running reg x. It's making sure that these models exist, right? Random force, gradient boost, logistical regression.
¶22It's making sure that everything is expected so that the script actually works. And you know, yes, these models can do more. They can validate their own work. And you see this when you run some of these state-of-the-art models. They automatically validate their work.
¶23They run scripts, they compile, they run llinters, they run tests, so on and so forth. This is no different. We're just putting the validation before the execution. We're saying how can we know for a fact this piece of code this unit of work is done and we build that as a script before anything runs. So this is us fighting against the second constraint of agentic coding.
¶24There's two constraints of agentic engineering. There's planning and reviewing the fusion harness with /autov valalidate lets us push past the current review constraint. Last week we broke down loop engineering and we focused on much more valuable concepts software developer life cycle ADWs and software factories. If you haven't checked that video out, definitely do that after you finish watching this one. That one went viral for a good reason.
¶25What we're doing here with the fusion harness is we're building on a single agent node. So the entire harness would be a single agent step. Literally once again scaling our compute to scale their impact. Instead of an individual agent running, we have a tight team of agents validating their work with each other, creating validation scripts before anything runs. As you can imagine, if this fails, the validator is going to loop back into the builder and say, "Hey, you messed this up.
¶26Correct it." We're adding agents into individual agent nodes to make up three value creators in the age of agents. It's engineers, its agents, and it's raw code. Again, check out last week's video to really understand the shift away from loop engineering because loop engineering is not encompassing enough. It's a terrible rebrand of the software developer life cycle. I hope you can see why owning your agent harness is so important.
¶27Let me just like say this very loudly and clearly again. Your tools directly limit what you believe is possible. When you use tools like the PI coding agent that are customizable and extensible by design, these mental false limits just fall away very quickly. You're not waiting for someone to deliver an update on some feature that's blocking you. Hint hint, cloud code, codeex, open code, so on and so forth.
¶28I mean this all the way down to like the fabric of the py coding agent. It's meant to be customized. This is the true value proposition of the PI coding agent and software that adapts to your will. There's a big ongoing debate, ongoing concern that we're all prompting all of our language models, all of our agents and we're giving away our company's data to these big AI labs, anthropic, open AI, Gemini, so on and so forth. This discussion of sovereign AI is going to become more and more important.
¶29We need to own our AI. We need to have local compute where we can run our language models knowing that our traces are ours. The most valuable piece of our work, the intelligence we're deploying, all the work we're putting into embedding our IP into agents is getting gobbled up by these AI labs. You know, make sure you like and subscribe cuz next week I'm going to address the question head on. Is anthropic and open AI stealing our data while we pay for it.
¶30That's an aside, a really important aside that I've been thinking about more and more and I want you to be very, very aware of as you and I engineer in the age of agents. Owning your agent harness is ultra important. Now that you understand this, now we're going to really scale this to the max. Okay, so let's throw a a harder, more realistic problem [music] at our fusion harness. Okay, so we're going to reset and now we're going to run FHS fusion harness state-of-the-art models.
¶31So let's imagine a concrete problem. I have an SQLite database inside of one of my users devices among thousands of users. SQLite is fantastic. There are over a trillion SQLite databases around the world. Knowing when to use SQLite versus Postgress versus some NoSQL solution can save you loads and loads of trouble.
¶32I need to insert over a million rows into an SQLite database across all my users. Am I going to sit down and just prompt this solution into a single agent? Insert all 1 million rows into this user into that user. Sure, we could. In this scenario, we absolutely could.
¶33But I'm thinking ahead. I know I'm going to have to solve this problem again and again and again for my business. And I don't want a one-off vibecoded solution. I want to solve every scenario in a fast memory efficient way. So what we're going to do here is use the fusion harness to solve this problem.
¶34I want the best state-of-the-art models near their maximum capability. We're using X high here on Clawed Fable 5 and X high on GPT 5.6 Soul. Okay. So we have an architect and a builder. And I want to know what is the best way to insert a million rows in a fast memory efficient way.
¶35So, of course, same flow / opinion. We'll paste in this prompt and fire it. Feel free to pause to read the exact prompt. We're just going to fire this off. And now we have two state-of-the-art models thinking through this problem.
¶36I want the opinions of the best compute available. As you can see here, I am paying for this directly through the PI coding agent. Both her agents are going to run here and put together a comprehensive answer for us. This is a slash opinion from two of the best models in the world right now. By the time you're watching this, you might have the new Gemini model.
¶37I actually don't think that model is going to be able to compete with Soul or Fable 5. We'll see in the stats. We'll see in the benchmarks. Regardless, it's going to be more compute available to you and I to build with and to combine our compute against. We don't need single individual winners.
¶38As you can see with the way things are really working out in the AI industry, there going to be many models. There going to be many players. And picking one is a mistake. Picking one model, picking one model provider, picking one agentic coding tool is a mistake. Here's a crazy stat.
¶39I'm building a new Asian harness nearly every single day. The Fusion harness is just one in a long line of custom agent harnesses I built out to solve specific problems to outperform any other out- of-the-box agent coding tool. Something to really think about if you're doing real production work when you're thinking about putting together your AI developer workflows that contain engineers code plus agents and your individual agent nodes. You can put any agent harness you need to or want to to get extraordinary results. Cloud Fable 5 done.
¶40Took about 90 seconds. Soul continuing to work. It looks like it's actually running some tests. It's actually kind of answering this question for us [laughter] before we even get to it. Fusion, of course, as you know, is coming next.
¶41A good workflow I like to use with the Fusion harness is, as you've seen, ask for an opinion, fuse the results of both opinion, and then you run the auto validate kind of build against it. Again, last week's video was so important for many reasons, but as you can see here, we're building out micro software developer life cycles, micro SDLC's. We're doing that inside of this harness. You can think of / opinion as like scout. You can think of /fusion as our unofficial planning step.
¶42And then, of course, autov validate is a twoin one. We're testing and we're building. Really curious to see how valuable its answer is in comparison to Fable. You can see it's already used 3x the tokens to generate this result. It's thinking a lot.
¶43We are in X high mode, so it's to be expected. But regardless, we're going to get the value of both models. There are engineers using fusion harnesses like this, using multiple models, getting multiple perspectives, and they're making better decisions. Let me be super clear about that. Engineers that can scale their compute to scale their impact in a functional, useful way, are making better decisions.
¶44This ties really closely into the CEO agent harness we built on the channel a couple months back. Again, a custom agent harness. We're customizing our experience to get specialized results. What is alpha? What is asymmetry?
¶45It's specialization. It's doing what others aren't yet. And so there we go. We finally got that response. We got some really, really insane speed ups from both model.
¶46Fable got its top speed up at uh 300x, but check this out. GPT 5.6 Soul found a 250 to 10,000X and another 400 to 2KX. Let's see if these models are bullshitting or if there's some true validity to what they're saying. These models are just thinking. They're reporting back in their own separate thought chains.
¶47Let's fuse the compute. Let's combine the context windows. Let's combine the intelligence. Okay, so fusion. And I'm going to paste this in here.
¶48This is going to be a longer one, but the idea here is that I'm having the models design one benchmark script that settles the bulk insert question for good, right? On two axes, speed and RAM. I'm kind of defining some parameters for them. This is an Astral UV single file script. Standard library only 1 million rows, identical schema generation, blah blah blah blah blah, right?
¶49Measure peak RAM. You get the idea. Pause the video here if you want to know the exact details. But here we go. They're off to the races.
¶50State-of-the-art models, the best compute money can get you, the best decision-making money can get you, the best engineering money can get you, minus actual engineers. Okay, [laughter] that still holds true. But I will say as a fact, the best engineers are using these models together. The best engineers are now like symbiotes with these state-of-the-art models. They don't make a decision without building, testing, validating, ideulating, planning with state-of-the-art models.
¶51One of the highest leverage areas for engineers now for agentic engineers is decisionmaking and it's planning. And guess what these models are great at? Uh, increasingly so. Even if they don't give you the right answer, they give you a very concrete answer you can use to increase your impact, to make better decisions, to build better software, to build better solutions for your users, for your business, for your customers. And so, they're off to it.
¶52This is going to chew up lots of thinking tokens. You can see here Fable is a lot more decisive with its building, with it planning. Soul taking its sweet time in X high mode. You can see the context window differences. You can see the cost differences.
¶53Fable already what is this 10 times more expensive than soul but also table 5 is already done here with its outputs. We're going to see what our builder puts together here. So very very simple idea taken to a reasonable level. Scale your computer to scale your impact. Great.
¶54I've said it a million times how you can use multiple models at the same time to solve problems. Great. You know this. You already know about delegation. You already know about spinning up multiple terminal windows.
¶55Everyone's using git work trees different coding agents. Blah blah blah. Cmox herder t-mucks. Like that's all great. What we really want to do here is combine the intelligence to make the best possible engineering decisions.
¶56This is where we deviate. This is where we push a little further than your classic sub aent delegation. Right? We're not handing off a task. We have two agents, really tight-knit team working together.
¶57We want their opinions. We want to fuse their results and then we can run auto validate loops that generate verification and building at the same time. And as these models continue to progress and as your harness engineering, your prompt engineering, your contact engineering continue to progress, all of this comes together to create some insane engineering results. And this is once again, I just want to mention this. This is just one node in your ADWs in your AI developer workflows.
¶58Bloom engineering is the wrong interpretation. It's the wrong mental model for understanding what you should be focused on right now. It's AI developer workflows. It's your software factory. Again, check out last week's video.
¶59I'm starting to get the feeling here that uh soul is going to create an insane result here. Hopefully, it's thinking is actually generating value. Some of these models just chew up thinking tokens and generate the same or worse outputs because they're overthinking it just like you or I might overthink something, you know, analysis paralysis. But here we go. SQL bulk benchmark builder OpenAI GPT 5.6.
¶60Here we go. It's running it. Uh this is going to be awesome. Using a lot of compute here to get this result. This is why a lot of the benchmarks, a lot of the results still say Fable is in the lead.
¶61The work it does, it just needs fewer tokens to ship. The simple way to put it, it's just smarter. Like a great principal engineer, they know when they don't have to do work. Knowing when not to work is just as important, maybe even more important, than knowing when to work. It's like when a senior engineer learns that deleting code is often better than adding code or modifying code.
¶62Anyways, skip ahead a little bit for you so I don't waste your time. There we go. GBD 5.6 took Soul wrapping up its response here after 8 minutes. So we have the total run for both of our agents. You can see here Soul was actually twice as expensive as Fable, but now Fable in fusion mode is going to combine the results.
¶63Our architect is going to look at both results and put them together. So it sees both source scripts. It's going to pull them together and it's going to fuse the result. So you can see there's the response from our agents. Looks like we have max tune gen with an insane 560x speed up for inserting rows over native SQL light autocommits.
¶64This is the real power of the fusion harness. It's really in this/ command. Opinion matters. Autov validate is a very very powerful pattern I highly recommend you check out for solving problems when you want to do less reviewing and when you trust your agents, right? Specifically when you have agents that understand your system, when you've created a specialized agent, a custom agent that understands your system better than others.
¶65One of the key leverage points there is, of course, the system prompt. You have to overwrite the system prompt to really guide your agent to powerful results. But even with the defaults, you can get really powerful results out of having a state-of-the-art model auto validate another state-of-the-art model. These capabilities will continue to increase. I know it sounds a little wacky.
¶66You know, some engineers rightfully so keep pointing to the error rates when you keep stacking up these agents. I think that's fine. that that is only the case when you haven't templated your engineering into your system and when you're not really putting in the effort to create great prompt context and harness engineering into your system. These out of the box agents are fallible. It is your actual work, your agent engineering in all these domains, right?
¶67All these subdomains, prompt, context, harness engineering that really make up the true value. You can see here the fusion agent is combining these results and it's going to use the best of both worlds. Two state-of-the-art models beats one. Two state-of-the-art models where you fuse the best of both worlds beats two. An auto validation loop using both agents with all their context.
¶68One to build the validation, the other to actually build the work. They iterate on each other beats just a build test workflow. Once again, scaling compute to scale your impact. And I really want to hit on this key theme. Your agent harness is the body that transforms compute into intelligence that works for you.
¶69So, whoever owns your agent harness owns your results. And this is why, you know, week after week I keep talking about the PI agent harness. I don't care if you use a PI agent harness. Build your own PI agent harness. Build a tool that does the same thing.
¶70Doesn't matter. The key is like you want to unblock yourself from being stuck on codecs, cloud code, open code. It is the customizability and extensible design of something like the pi coding agent that truly differentiates it and truly differentiates your results. The workflows that we've looked through here are just a few, right? There are many other patterns you can build into your fusion harness.
¶71Couple ideas just to throw at you. Debate. You can build a slash debate. Pass in a number. Pass in a prompt.
¶72Pass in a total debate rounds. So n 10 times, five times, three times. Have your agents debate an idea or a concept or an architecture. You can do like / parallel. And this should just execute both of your agents at the same time versus opinion.
¶73They're just responding. Opinion doesn't have right or edit tools. There's something like uh you know a coordinate. You give your agents a task. They talk back and forth about the solution and then they actually build it.
¶74So this is all in the realm of harness engineering and in that there's prompt engineering, there's context engineering. All of it fits under the subdomain of agentic engineering, right? Building valuable software with agents, with code, and with engineers as the three units of value creation. Our agent finish here. Let's see what the fuse result looks like.
¶75What's the best strategy? So the speed winner setbased CTE. This is a combination of builder and architect solution to get an a,000x speed up. Love that. Here is our memory winner wallt tuned generation.
¶76Not going to go into too much of the details here. If you're a database nerd, uh you'll be able to look at these results. I'll make sure to save this in this codebase so you can take a look at these. But the consensus divergence, both the models independently chose subprocess per strategy isolations, one deterministic data set, a bunch of stuff. So they matched up here.
¶77They diverged here. This is really interesting. The builder solution GBD 5.6 soul was simpler and faster, but the three column schema was simpler from the architect. single run order over median of two shuffled simpler that was from the architect. Okay, so there's just a bunch of stuff here where uh one idea was pulled from the architect, one idea was pulled from the builder, one idea was pulled from the architect, they both got the same idea, so on and so forth.
¶78These divergences is what makes value. It's not the fact that you have 10 John's working on your team. It's the fact that you have a John, a Sally, Tim, you know, insert the analogy I'm trying to make, right? It's the range of perspectives and the range of ideas on your team of engineers on your team of agents that gives you the real value proposition. Okay.
¶79So this is why I say don't choose compute combine compute thinking ands not ors. And so you know the last prompt here is the of course autov validate. It looks like our agents kind of already did run this to like really build this out. I'll kick this off. This runs the exact same workflow.
¶80You know our validator agent is going to read the work. It's going to create a comprehensive gate file that our builder must run when it finishes. Again, this is built into the fabric of the fusion harness. Because we own the Asian harness, we can make it so that the builder cannot edit the gate file. And then if something goes wrong, it'll loop and our agents will coordinate together.
¶81This is a team. This is not subent delegation. We have agents working together to deliver you better results. So, I'm going to leave this PI coding agent fusion harness linked in the description if you're interested in how to combine models to outperform either. Just like agents plus code beats agents alone and agents plus code beats code alone, multiple agents that fuse their intelligence beats the agents standalone.
¶82The industry is coming around to this idea once again. We're renaming it. I don't care about the name. I don't care who's writing a blog post about it. I want you to have these ideas so that you can scale your compute, to scale your impact, so that you can win in the age of agents.
¶83This is another multi-agent orchestration pattern and it's another node inside of your ADWs or AI developer workflows. For those that watched last week's video that went pretty viral for a deep tech channel like this, we dismantleed loop engineering and discussed what we should really be focusing our agentic engineering on. And the key topics there is the software developer life cycle, ADWs, and software factories. Check out that video if you haven't because the fusion harness fits perfectly inside of a single agent node and it's all making up what's possible. It's one piece of the puzzle of the frontier of Agentic.
¶84I'm going to keep pushing this harness forward as I have several of my other custom agent harnesses because the value proposition is there. We've talked about the verifier agent harness. We've talked about the CEO agent harness, multi-team agent harnesses. The frontier is right in front of us and we're just getting started. Next week I'm going to dive into a very very important spicy topic is really critical for you and I as engineers building companies building software building IP that we want to defend.
¶85We're going to ask and answer the question as long as nothing else you know more important comes up. This landscape is so crazy. I want to dial into this question is anthropic and open AI stealing our data while we pay for it. Stealing our IP, our valuable prompts, our valuable business workflows, are they stealing it while we pay them for it? Of course, this leads directly into the conversation about Chinese models, about US openweight models, so on and so forth.
¶86That's the topic for next week. If you made it this far into the video, huge thanks to you. Make sure you like, make sure you subscribe, all that good stuff. Make sure you comment, let the algorithm know you're interested in true agentic engineering on this channel. We don't just clickbait the news.
¶87We think, plan, and build to progress our engineering in the age of agents. Our Fable 5 validator just finished this [music] nice massive validation script so that once it runs, we know for a fact the code is going to work. There is the initial pass fail [music] and now our builder is getting to work actually creating the script. I'm not going to hold you here for this. I'll make sure to add the result into the codebase.
¶88Again, link in the description for you. We have a lot of really important concepts to discuss on the channel coming up. So, be sure to like, subscribe, comment, all that good stuff so the algorithm doesn't miss you. You know where to find me every single Monday. Stay focused and keep building.