Real-time observability, guardrail layer for AI agents proposed

This is a solid architectural discussion hitting a real pain point — existing APM tools are fundamentally not designed for non-deterministic, nested agentic systems where a single user request can spawn dozens of sub-agent calls.

• –The Go proxy layer for token/latency tracing per sub-agent (not just per-app) addresses a genuine gap: most teams today have zero cost attribution below the application level
• –Neo4j for tracking agent call graphs is a natural fit — graph traversal maps well to nested tool call hierarchies and cross-session attack vector analysis
• –The PII/prompt injection guard middleware is the hardest part: semantic redaction at the proxy level without adding latency is unsolved at scale, and the post wisely flags this tension
• –Shadow AI discovery via cloud audit log scanning is an underrated enterprise need — most orgs have no idea which models are being called by which services