Benchmark reveals AI agents ignore long policy documents

Long-context windows might allow models to "read" large policy documents, but they completely fail at ensuring agents actually follow those rules when distracted by complex, multi-step tool use. The failure patterns are highly consistent: agents frequently let a plausible user request override a strict standing policy, and often perform required compliance checks but then inexplicably act against the results. These findings suggest that for enterprise deployments, critical policies cannot just be prompted; they must be enforced deterministically through tool guardrails.