Writing

Three axes, more than one lever

Cost, quality, and latency move together in an agent deployment. A monolith gives you one model to move them with. A role based harness lets you choose per component, but only if you can see which component moved which axis.

Every instrument favors what it grew up next to

An eval harness built alongside one architecture quietly encodes that architecture as normal. Most of the biases I found in my own measurements pointed the same direction, and every one favored the thing I already had.

"Because LLMs are non-deterministic" is not an answer

No-code agent builders are great for demos and the happy path. When someone asks why the system did what it did, that's where the real architecture question starts.