What we've learned
from deploying AI in
the real world.

A great agent works around the problems in its tooling and finishes the task anyway. That is correct behaviour, and it hides every one of those problems from us. So we gave our agent a tool to file a ticket: about six a day, thirty-seven over six business days, most of them real issues we were not thinking about. Plus the four other tools we run alongside it, and why evals come last.