Developer shares three critical bugs found after deploying a multi-agent AI system

A software developer built a production-grade multi-agent AI system using Python, LangGraph, Redis, and ChromaDB, featuring a supervisor agent, persistent memory, human-in-the-loop approvals, and full observability. A flaw in the escalation logic caused a single task to generate over 5,000 duplicate database rows, which was ultimately resolved by adding a partial unique index at the database level rather than relying solely on application logic. A second bug involved a silent fallback in the checkpointer initialization that allowed the system to run without state persistence, appearing functional while failing at a core task; the fix was to replace the fallback with a hard startup error. A third issue caused the replay feature to compare incorrect states because the frontend was sending an OpenTelemetry span ID instead of a LangGraph checkpoint ID, with a silent fallback masking the mismatch. Across all three bugs, the developer noted a recurring theme: silent fallbacks and absent database-level constraints allowed failures to go undetected, making systems appear healthy while producing wrong or corrupted results.