An AI agent that checks the work of the studio's other AI agents. It doesn't fix anything
or make decisions. It reports what it finds, and the studio's owner decides what happens next.
1. The question
The studio runs several small apps on one computer. Some of them kept stopping, and nobody knew why.
Another AI agent, the Platform Engineer, offered an explanation:
The Platform Engineer's explanation
Apps started by an agent run inside the Claude app's group of processes and can't leave it, so
they stop whenever the Claude app closes, even ones started on their own. It was about to build
a workaround.
Before anyone builds a fix, it's worth checking whether the explanation is true. That's this agent's job.
2. First, test the checker
The check uses a small tool that asks Windows which group a process belongs to, and how that group is
set up. If the tool gives wrong answers, everything after it is wrong too.
So the tool gets a question with a known answer first: put a test process in a brand-new group with one
setting turned on, and see whether the tool reads that setting back.
Not run yet.
3. The evidence
Now the real questions, measured directly on the studio's computer rather than taken from anyone's account.
Not run yet.
4. What did the check actually show?
A report keeps what was measured apart from what's only a guess. Sort each statement.
5. The report
Half right, half wrong
Right: the Claude app's group can't be left.
Wrong: an app started on its own isn't in that group at all. And the apps that kept
running overnight were in a different group, Python's own, which closing the Claude app doesn't touch.
Still unknown: why the apps stopped. Nothing measured so far shows the cause.
What happened next
The Platform Engineer withdrew its explanation and dropped the workaround before it shipped. Apps started
through the studio's launcher now record their group when they start, so the next time one stops,
there's evidence to look at.
This agent also withdrew one of its own earlier readings from 28 September. It had made the same
mistake: it took Python's group for the Claude app's.
This is a real check from 6 October 2026. The outputs are what the checks printed that day, with
file locations removed.