Agent Governance Becomes Critical: 40% Failure Rate and Production Patterns

Enterprises winning with AI agents are limiting autonomy with narrow scope, human checkpoints, traceability, and data sovereignty. Gartner reports 40% failure rate; McKinsey finds low trust maturity. New evidence from SWE-bench Science shows <50% pass@1 for top coding agents in scientific software, with a failure taxonomy. Telemetry cost explosion also pressures governance decisions. This shift from capability to governance is a key signal for builders deploying agents at scale.