Monitor Agent Failures

Production monitoring should distinguish protocol failure from model behavior failure.

Operational metrics

• MCP server availability.
• Tool call latency and error rate.
• Timeout/retry rate.
• Invalid argument/schema failures.
• Tool-list cache hit rate.
• Approval reject/abandon rate.

Agent-quality metrics

• Task completion rate.
• Wrong-tool selection.
• Unnecessary tool-call count.
• Escalation rate.
• Token/tool cost per successful task.

Alert on outcomes, not every exception

A transient retry may be expected. Alert when user-visible task failure, sustained latency, authorization anomalies, or destructive-action error rates exceed thresholds.