| Captain’s Log, episode 1 |
17:22 |
Rita Rivera and Skippy the Magnificent. |
I handed four AI agents a written plan and went to bed. In the morning one of them had
refused to do the job.
Not broken. Not confused. It had read the plan, tried to build it, and left a note
explaining that what I asked for was impossible. The plan said you could tell whether an
address is in Canada or the United States by how far north it is.
Windsor, Ontario is south of Detroit, Michigan.
This episode walks through what four agents found overnight with nobody watching, working
on a Dynamics 365 Field Service tool that tells a repair crew whether it is safe to drive
into a storm. Three of the four findings were the kind that gets somebody hurt: a forecast
that returned a confident all clear on hours nobody had looked at, a rule that compared
miles per hour against degrees Fahrenheit and threw away a real wind warning, and an alert
that argued with itself in front of the one person who had to act on it.
Then the part that matters more than any of it. The next day, with 169 tests passing, the
software was quietly pulling weather for the wrong city. Not one check caught it. I found
it in four seconds by asking the real weather service what it thought.
A test fixture is captured from a request. It proves your parsing. It cannot prove the
request was right. The loop verifies coherence. It cannot verify contact.
Full transcript, verbatim agent quotes, and the numbers that do not flatter us:
https://nukasoft.ai/blog/engineering-notes/minutes-of-a-meeting-that-actually-happened/
