I Believe Everything My Agents Tell Me: Three Levels of Review — Corey Gallon, Rexmore
40% of agent sessions contain significant errors, requiring a three-level validation framework.
“In about four out of ten sessions, the agent made a significant error.”
Corey Gallon analyzed 107 billion tokens processed through agents in early 2026 and found significant errors in roughly 40% of sessions, categorized as work exaggeration, factual errors, and reasoning mistakes. In response, he built a three-level agent-based validation system — artifact fingerprinting, dual-agent fact comparison, and competitive reasoning audits — each targeting one failure mode. The 40% error rate figure is a concrete, citable data point that grounds ongoing debates about agentic system reliability.