What shipped, and what the numbers said

Six weeks after the rebuild, the numbers had moved enough to be worth reporting — and one of them had not moved at all, which is the part worth reading closely.

That last line is the honest one. The agentic branch — the interesting part, the part that justified the project in the founder’s mind — moved four points. The boring routed workflows delivered nearly all the value. The team had been most excited about the least productive third of the system.

What they would tell you they would do differently: read 300 tickets before writing any code rather than after, ship the observing-and-proposing stage first instead of going straight to drafting, and put the cost-per-ticket chart up on day one rather than after it had cost them $80 to notice a broken carrier.

And one thing they got right by accident. Because refunds always required a person, the injection in week three was a curiosity rather than an incident. Nobody had anticipated that specific attack. They had simply decided, early and for unrelated reasons, that money leaving the company needed a human. That decision paid for itself the first time it was tested — which is the usual way a permission boundary earns its keep.