What’s the hardest part of building across multiple AI providers?

We’re launching LotaGate soon, and one of the core problems we’ve been working on is fragmentation across the AI development stack.

Integrating a single provider is usually straightforward. Complexity starts to grow when teams work across multiple models and providers — different APIs, SDKs, authentication methods, streaming formats, error behavior, routing, observability, and operational tooling.

LotaGate is being built as a unified infrastructure layer around that problem, with an AI Gateway, SDK, Agent SDK, CLI, Desktop App, and Web Platform.

Before launch, I’d love to hear from other developers:

What creates the most friction for you when building across multiple AI providers?

Is it API compatibility, model switching, routing and fallback, cost visibility, reliability, agent workflows, observability — or something else?

I’m especially interested in real production pain points and workflows that existing tooling still does not handle well.