Incident summary
A logistics platform allowed automation to rewrite routing policies in production when congestion thresholds were crossed.
A bad telemetry batch falsely marked healthy lanes as degraded. Repeated policy rewrites then pushed traffic into already saturated routes.
What this teaches
- observability is not enough without control authority boundaries,
- confidence scores must gate high-impact actions,
- rollback plans must be prevalidated, not improvised.
Practical controls
- Confidence gate regional/global policy changes.
- Require human approval for cross-region rewrites.
- Rate-limit live policy write frequency.
- Keep one-click known-good rollback bundles.
Short note: bldrAgent Workflow Board can enforce these approvals, but the same approach works with any policy-aware workflow.
