We're the ones who picked

We built a few things and we keep them running ourselves. That is where anything new goes first.

Only what holds up there for a few months comes up in a client conversation. What does not, does not.

Old code makes "let us rewrite it" the easy first sentence. The trouble is that a rewrite spends real money and often leaves the screen your users see exactly where it was.

When we do say replace it, we point at what is blocked right now. If we cannot point at anything, what is there stays.

Connecting an outside platform, we do not read the docs and declare it all supported. There are platforms where listing a product works fine and returns have no endpoint at all.

So what works and what does not get split apart, and where something does not work the button never appears. Better to lose the button than to have it fail after the click.

The unit these platforms count calls in differs too. Naver counts per advertiser, Meta counts per use case. Put them behind one counter and the moment one is throttled, all of them are.

The editing agent we built has five brakes on it. How much it can spend before it halts, how many seconds one call gets before it is cut, both sitting there as numbers.

It ships switched off, and a person is the one who switches it on.

One more thing. The unit AI counts text in fills up about three times faster in Korean. Leave the English defaults where they are and the savings never arrive. Those thresholds were recalculated in Korean.