Something that decides what to handle itself and what to send out to somebody else is not a router.

It is a manager.

That is the whole of what follows, and it is a bigger claim than anything on Experiential Labs' website.

What they built

Experiential Labs is building an open source AI gateway. It sits between an organisation and the models it uses. Whoever holds the AI budget can control who reaches which model, and the traffic passing through can be turned into models the customer owns.

That is the product as described. The interesting part is the position it occupies, which is the place where every request gets looked at and a decision gets made about where it should go.

The founders are Kion Fallah, chief executive, previously a staff research scientist at Waabi leading the mixed reality simulation team, with a machine learning PhD from Georgia Tech, and Silen Naihin, chief technology officer, who built and grew AutoGPT to 160,000 GitHub stars, did AI for science at the Department of Energy, and previously cofounded Stackwise. It is in the Summer 2026 batch.

Who else is in this field

Gateways and routers are a busy category. OpenRouter put one interface in front of hundreds of models. Portkey, LiteLLM, Helicone and Cloudflare AI Gateway all sit in the same slot in the stack. Kong and Apigee did this for ordinary APIs long before models existed. Ollama and LM Studio made running a model on your own hardware easy. OpenPipe and Predibase move work onto smaller models you control.

Every one of them is treated as plumbing. Plumbing carries things. It does not decide.

The point: we are hiring an AI to manage the other AIs

This breed of company does not simply have an AI. It has an AI that knows what to think for itself, cheaply and locally and safely and privately, and what to hand out to somebody else.

Those are two genuinely different judgements, made fresh on every request, and making them well is a skill rather than a configuration.

The handing out gets more interesting than it first sounds. Today the question is roughly whether to send a request outward at all. Before long it becomes a question of where: which outside model is best at this particular job, at this price, with this much delay, under these rules about where the data is allowed to travel.

We are hiring an AI manager to manage all the AIs. It decides what it keeps and what it delegates, and it earns its keep either by saving money or by automatically picking a better specialist than the human would have picked.

Which is exactly why it cuts into human work so easily. It does not have to be smarter than anybody. It has to make a routing judgement thousands of times a day, slightly better and very much faster than a person choosing a model out of habit. That is the shape of work that machines take from people: not the brilliant decision made once, but the ordinary one made constantly.

And it is the right thing for an AI to be doing. Most of this field is building intelligence. Comparatively few are building the judgement about which intelligence to use, even though that judgement is now in front of every single request anybody makes.

Which means it ends up facing the human

Here is where the argument goes, and it is a prediction rather than the company's stated plan.

Something that decides what kind of brain each request deserves does not stay behind the scenes. It ends up owning the layer the human actually touches. Not the model, and not the device. The thing in front of both that works out what this person is trying to do, and what sort of ability the job actually needs.

So this company is not the AI. It is one of the first to refocus on the human interface, on what it takes to be the first layer a person interacts with.

That reframes the interface argument everybody is currently having. The debate is about form factor: voice or text, screen or glasses, chat or agent, and which of them wins.

The better question is not which form factor wins. It is which occasion deserves which form factor.

Nobody wants to speak aloud in an open office. Nobody wants to type a long brief while walking to a meeting. Some questions deserve an instant local answer and some deserve a slow careful one from the best model money can reach, and the person asking almost never knows which is which beforehand.

That is not a design decision made once and shipped. It is a pattern learned by watching what people actually do, in which situations, and which responses they accept.

Which a company optimising the experience collects as a matter of course. Not as surveillance, but as the natural residue of trying to make each interaction land properly. It is presumably why they are called Experiential Labs rather than Gateway Labs.

So the long version of this bet is that the way humans give instructions to machines is about to change again, and that the change happens at the layer deciding what to route where, rather than at the model or at the device everybody is currently arguing about.

Why it is bigger than it sounds

The short description is an AI gateway, which sounds like plumbing.

  • It is a manager, not a router. Deciding what to think locally and what to delegate, and eventually where to delegate it, is a judgement rather than a pipe, and it sits in front of every request a company makes.
  • It does not need to be smarter than anyone. It needs to choose better than a person picking a model by habit, thousands of times a day, which is precisely the work machines take over easily.
  • Whatever decides which intelligence a request deserves ends up facing the human. And the useful interface question stops being which form factor wins and becomes which occasion deserves which form factor, which is learned from behaviour rather than designed.

What to watch

One thing decides whether the manager idea is real: whether the routing is actually a judgement. If the gateway is weighing cost, delay, privacy and fitness for the specific task, and getting it right more often than the engineer would have, then there is a manager in there. If it routes on a static rule somebody configured once, it is a proxy with a dashboard, and the whole argument above collapses.

The second is whether they can tell you which occasions suit which kind of response. A company that has been watching real traffic for a year should be able to say something about that which nobody else can, and being able to say it is the first evidence of the interface position.

The third is the question to put to them directly. Do they see themselves as infrastructure sitting behind other people's products, or as the layer the human addresses. Those are very different companies, and only one of them is the one described here.