Persistent Failure Modes in Tool Augmented AI Systems: Constraint Use and Trajectory Intervention, Evidence Review and Candidate Monitoring Protocols

This paper reviews recurring failures in tool-augmented AI systems through two analytical axes: retrievable but unused applicable constraints, and intervention outcomes that vary with timing, format, and execution state. Version 2.0 distinguishes harmful from beneficial sensitivity and does not infer task-specific capability from general benchmark performance. The evidence review compares harness, handoff, and communication studies with explicit task, information, resource, and denominator differences. Six monitoring candidates require independent future endpoints and account for unresolved and terminal failures. No new agent experiment is reported. Joint-regime and transfer-compression hypotheses remain unvalidated, and the earlier general claim that interruption is harmful is not retained.