TanStack's AI kit hedges against model churn

Model releases slipped from June into July, and I stopped betting my code on any single provider.

The model market moves weekly now. Rewriting an integration every time a better model or a new agent harness appears costs real time you do not get back.

Hard-code calls to one provider, and every swap means touching every file that calls it. TanStack ships a fix for that in a library called ai, and a tool called manifest works the same way, per TanStack's GitHub repo.

Both are type-safe and provider-agnostic. The TanStack kit streams responses and runs tools and agents across OpenAI, Anthropic, and Gemini. Swap the provider, or the harness running underneath it, and the code calling it does not change.

That is the trade worth making right now. The model you picked this month is already replaceable. The abstraction layer and the specs you wrote around it are not going anywhere, and that is where the real value sits. If swapping models ever costs more than a config change, I will take that as proof the abstraction failed.