Failure Analysis of Stacked Deterministic and LLM Agent Gates

Testing 1,119 agent actions reveals that stacking LLM judges provides minimal error reduction, with multiple judges composing to only 1.2 to 1.4 effective layers. Combining a deterministic rule layer with one LLM judge yields only approximately two layers of protection.