Compiled query execution is a solved problem — HyPer proved it, SingleStore productionized it. The open question isn't whether to compile SQL, it's what IR to build the compiler on. Every major compiled database goes straight to LLVM IR. PhoebeDB doesn't. This post explains why, and what three custom IR layers buy you that LLVM IR alone cannot.

Every team building a serious AI agent system hits the same wall. PostgreSQL for state, Pinecone for vector memory, Redis for events, Kafka for streaming — five operational surfaces, five billing accounts, five places where something can go wrong. PhoebeDB was designed from the beginning to solve this fragmentation with cross-type atomicity, snapshot-consistent reads, and causal replay in a single kernel.