arXiv 2607.05775 clusters 19 benchmarks into six LLM agent failure modes

The synthesis groups documented errors from tool invocation papers, planning audits, and multi-agent evaluations into a pipeline taxonomy. Tool parameter errors and constraint violations appear in single-turn settings; these compound with context length, producing nonlinear drops in end-to-end success on tasks exceeding eight steps. Scaffolding additions improve isolated sub-tasks yet fail to raise aggregate completion rates across WebArena, AgentBench, and ToolBench traces.

Data show short-horizon navigation and single-function calls reach 70-85 percent accuracy while full-horizon trajectories fall below 30 percent once cumulative context exceeds 8k tokens. Measurement validity gaps arise when benchmarks reward partial credit or permit oracle tool lists, inflating scores that do not transfer to open environments. Safety failures concentrate under underspecified goals rather than explicit jailbreaks.

Operationally, current agent deployments remain gated by horizon length and verification overhead. Production systems therefore retain human checkpoints or restrict scope to sub-task loops where per-step reliability exceeds 90 percent. The taxonomy indicates that further leaderboard gains on narrow slices will not close the gap to reliable autonomous operation without new verification primitives.