I’ve been working with Generative AI systems across startups and enterprise teams, and I keep seeing the same architectural gaps: missing patterns around retries, observability, workflows, and safety.

This blog outlines 8 major areas I think we haven’t solved yet: token-level tracing, agentic idempotency, AI state snapshotting, embedding versioning, etc.

Curious if others building in this space are seeing the same gaps or have already found good solutions.

Would love your feedback and discussion.