The Verifier Tax: Horizon-Dependent Safety–Success Tradeoffs in Tool-Using LLM Agents [R]

We recently presented a paper at ACM CAIS 2026 on safety evaluation for tool-using LLM agents. The core issue is that task completion alone can be misleading: an agent may complete a task while violating a safety or policy constraint. We separate outcomes into safe success , unsafe success , and failure , and study how verification changes this tradeoff. We evaluate this using τ-bench / Tau-bench tool-use scenarios and propose a two-tier verification architecture : deterministic policy/tool checks first, fo