LLM Agent Failures: A Unified Taxonomy of Limitations.

While Large Language Model (LLM) agents are increasingly evaluated on their ability to perform complex tasks involving tool use, planning, and multi-agent coordination, benchmark gains often mask persistent and recurring failure modes. This research provides a comprehensive synthesis of 27 papers published between 2023 and 2026, integrating findings from 19 distinct benchmarks to create a cross-cutting taxonomy of agent limitations. The synthesis identifies six primary clusters of failures. These include errors in tool invocation and parameter handling, breakdowns in planning and constraint satisfaction, degradation of performance over long horizons due to context accumulation, and failures in coordinating with multiple agents. Additionally, the taxonomy highlights safety and security vulnerabilities under adversarial or underspecified conditions, and issues related to the validity of measurement itself. The study reveals that these failures tend to compound nonlinearly as task length increases. It also notes that strong performance on individual sub-tasks does not reliably translate into end-to-end success, and that adding more scaffolding does not consistently improve reliability. Despite these challenges, significant progress has been observed in single-turn tool use, short-horizon web navigation, and narrowly defined coding tasks, suggesting that while broad reliability remains elusive, specific capabilities are advancing.