What the system caught
What happens after an idea becomes a build? These five records show what the agents built, what the checkers caught, and what changed. They show the engineering work behind the tools—not a measured increase in sales or marketing performance. Where no screenshot survived, the page says so instead of drawing one.
Misfit Hollow · review and repair
One version of Misfit Hollow, frozen so it couldn’t change mid-review. Provider names are left out. Anything blocked or put off stays amber.
Ten agents were requested. Seven could run and three were blocked, and none of the blocked ones launched.
Seven agents reviewed the app without seeing each other’s work. Two coordinators organized them but didn’t review the product. Together they found 18 problems or missing tests.
The seven reports became one list. A finding made the list only if it could be reproduced and pointed to a file and line, not because several agents agreed.
Four repair agents (server, interface, domain, and packaging and tests) changed nine files between them without touching each other’s, and handed back fixes rather than installing anything.
86 of 86 Python checks and 49 of 49 browser checks passed on the combined version.
A separate reviewer checked only the repairs and passed them, with limits. Two risks were put off on purpose. The fixed version passed its checks but wasn’t installed in this run.
Five runs
Each one keeps what was proven, what wasn’t, and an empty space where a real screenshot doesn’t exist yet.
Misfit Hollow · five builders
Five agents each built the app on their own, without seeing each other’s work, while a coordinator organized them without writing code. All five versions were kept, and one became the base for the next round.
ProvenFive working versions, one coordinator, about 23 minutes in the build queue. It doesn’t show that five builders beat one.
Misfit Hollow · mixed review
Seven reviewers, working separately, found that the app held a town-wide lock while it sent data, so one slow connection could stall everyone. The repair let go of the lock before sending.
Proven18 problems found, 16 fixed and 2 put off on purpose; 86 of 86 Python and 49 of 49 browser checks passed. This run stopped before installing.
PDF Workbench · fixing an app another agent wrote
A live test worked through editing, moving, deleting, undo and redo, reopening, saving a form and exporting, on made-up documents. The test tool gave up waiting for the download, but the file was there and opened with the right content.
LimitRotated or cropped pages, phones and multiple tabs weren’t all re-tested live.
Composer’s board · recovery
When the builder got stuck, the coordinator replayed only the edits it had already made, instead of launching a second builder. The test copy and the installed copy each passed 119 tests. When the board can’t read its data, it says UNAVAILABLE instead of showing an empty board.
MediaVerified screenshot pending. No mock has been substituted.
Composer · guarded install
A two-file fix, plus a test that stops it coming back, closed the leak. Composer passed 164 tests, and it still says plainly when it can’t read its data.
LimitComposer’s code started outside the Apartment; it wasn’t built there from the beginning.
Across all five
Our rule
No screenshot, no substitute mock. If a real capture doesn’t exist yet, the space stays empty until one does.