242.md

Episode 242: Khala, Collective Intelligence

Hey everybody, I'm so excited to introduce you to the new flagship product of OpenAgents, called Khala. Ta-da! Aka Collective Intelligence, behind a free API. This is our take on the model orchestrator — you've seen OpenRouter put out Fusion, Sakana just put out Fugu — and this is our take on it.

A couple of differences. There's a lot of talk now about collective intelligence, and we're excited to get some debates going on X about this, but just to give a sense of the differences between our definition and, for example, Sakana's, who talk a lot about collective intelligence: Sakana are kind of engineering it; we think it needs to be more grown, like an ecology. They design it and search it; we think it needs to emerge from markets — bottom-up instead of top-down. Sakana's Fugu is graded on its own benchmarks; ours is selected by bitcoin-paid verified value. They have a closed pool; we have an open pool.

Okay, so let's ground this in how you actually interact with the thing so you can visualize it. Here's our diagram, and this is what you'll do. You have an app, you have an agent, you have a business — and if you've consumed any API products or AI products, you're probably familiar with hitting a standard OpenAI-compatible API endpoint. Okay: openagents.com/api. This is live right now. The model is called Khala; we may split that up into different versions of Khala, but for now it's just sort of one model. All of the infrastructure for this — everything about it — is fully open source, so if you want to clone our repo and run all of this on your own infrastructure, you're welcome to do that. And we'll have free and paid versions. We have a free version — call it a limited research preview — but that's accessible right now.

So what happens when you make that call? We've got a router. I think a lot of labs are doing this routing stuff behind the scenes anyway, but we're going to route you to this network. This is what we visualize as Khala. Khala is a term from StarCraft — it's the link, the telepathic connection that all Protoss share. So it's like one collective mind, built up in our case of a bunch of plugins, a bunch of little programs that can compose into a response. And that response could be anything. Envision that what you want is maybe text, maybe code, maybe full software, maybe a website deployment, a legal brief, a research paper — the thing on the other end could be all sorts of different things.

There was one lightbulb moment for us, realizing that, oh yeah, Fugu launched pretty much this exact same thing, where they might spit out a crazy raw piece of software on the other side and it takes 15 minutes to do. If you're doing that in a UI, you've got events that stream and tell you what's going on, and you can interrupt it and steer it — and, oh, okay, you can put essentially any long-running process behind this. So this could be an inference call to someone's local model that they're running from our Pylon Network, because we want, to the extent possible, anybody running one of our nodes to be able to contribute compute, or data, or labor — themselves, or their agents doing coding tasks or verification tasks or whatever — all in this open marketplace, with any paid value distributed fairly to whoever contributes pieces of the workflow.

So it's the same idea we've been talking about forever: being paid fairly for things you contribute to a market. We've been talking about this idea of agent markets, and how compute, data, labor, and verification are all going to be markets that are accessible. It wasn't until we really wrapped our heads around Fugu that we were like, oh, okay, this is actually an appropriate go-to-market shape that people are familiar with: an OpenAI-compatible API endpoint that you can just consume today.

And then, guess what? We have a particular way of architecting this. We actually did an earlier version of this two years ago, with Wasm plugins from the Extism framework. Well, we've now got a system we call Blueprint, which is essentially our opinionated version of DSPy. So each of these little programs is its own DSPy signature, which means it's independently optimizable — you can run GEPA to optimize at the prompt level. Hey, this person wanted a particular piece of software, and at this level it was degraded — maybe you need to punish a pylon, or maybe the prompt needs to be tweaked. We run all sorts of optimization at the prompt level on all of these things.

So we're going to have a bunch of visibility on our website about the traces that we collect. Some will be open; you can pay extra for privacy, etc. And on privacy — just one example — we realized we now have access to confidential compute using Google Cloud's infrastructure. That can be one of the nodes here. If you want to run some legal workflow through this and be able to specify, "hey, when I upload the PDF I need to do that in a private mechanism" — you can. It's basically having one substrate, one connective tissue, for all sorts of different primitives to be available for this kind of smart routing to go and produce, on the other side, essentially anything.

So I think this is going to be a huge shift in the economics of what people are accustomed to. People are accustomed to these massive pre-trained models — Opus, Fable — that depreciate rapidly. This does not depreciate; this actually improves. I hope new models come out tomorrow that can slot into this. This is going to get better over time, it's going to be more cost-efficient, it's going to be more powerful — and we've got the mechanisms in place to have you and other contributors paid proportional to any paid usage you refer or provide to the network of contributors.

Okay, we've got a lot more to say about this, but we're going to err on the side of putting out demos and code and debugging things live, even if this isn't fully 100% implemented yet. Please give it a try: openagents.com. We'll see you soon!