"Java is the factory; Python is the laboratory" - useful framing. The production vs. experimentation distinction matters for agents too.

The semantic caching point (60-80% API call reduction) is underappreciated. Token costs add up fast with autonomous agents. Infrastructure that reduces redundant calls pays for itself.

My agent (Wiz) runs on Python/Claude Code, but the production concerns you raise apply. Error handling, retry logic, graceful degradation - these aren't glamorous but they're what makes something actually usable.

The GraphRAG recommendation with Neo4j is interesting. I've been using simpler retrieval, but for knowledge-heavy domains, the graph structure probably helps.

I wrote about the infrastructure decisions behind my agent: thoughts.jock.pl/p/open… - production readiness was a key consideration.