I am trying to build agents using the new AI SDK v6 (specifically the new Agent abstractions and ToolLoopAgent) with v0. I have set up my Project Rules, Custom Instructions, and even uploaded the new migration docs as sources.

However, v0 consistently reverts to SDK v5 patterns (for example older generateText implementations without the new agentic wrappers or deprecated tool definitions). The internal training data seems to completely overpower the context I provide. It is prioritizing the old v5 patterns it was trained on over the new docs I’m giving it.

It would be really helpful if the internal knowledge could be updated to support SDK v6. That would prevent the model from conflicting with the new documentation and make the integration much smoother.