“A 90% accurate process repeated 5 times is (0.90.90.90.90.9), or 0.59, 59% accurate.”
Honeycomb shipped an LLM query builder and found latency of two to 15+ seconds per answer. Customers with 5,000-field schemas blow past the context window. People are already using prompt injection to try to pull other customers’ data. Every agent demo that chains five calls together is quietly shipping a coin flip.