When Errors Become Narratives: A Longitudinal Taxonomy of Silent Failures in a Production LLM Agent Runtime

A study of a production LLM agent runtime identified a five-class taxonomy of errors, with 70% of silent failures detected by human observation, not tests or audits^[1].

The study, conducted on a personal-assistant agent runtime in continuous production since March 2026, analyzed 22 incidents with full root-cause postmortems over eight weeks. The runtime had 40 scheduled jobs, 8 LLM providers, a tool-governance proxy, and a knowledge-base memory plane. The taxonomy includes environment and platform quirks, design-assumption mismatches, error swallowing and dilution, chained hallucination and fabrication, and operational omission and forensic blind spots. Chained hallucination and fabrication, unique to LLM systems, is considered the most dangerous type of error. Meanwhile, a separate study on large language model (LLM) agent communication protocols developed a technical taxonomy to classify and analyze these protocols, comprising five dimensions: counterparty, payload, interaction state, discovery mechanism, and schema flexibility^[2]. The findings highlight the importance of robust communication protocols in distributed agent networks, with all sampled agent-to-agent protocols combining hybrid payloads with session-state persistence. Most protocols support multiple predefined schemas, and two negotiate schemas at runtime.