Agent builders are rediscovering loop control from scratch, one stuck agent at a time. Psychology has studied the difference between a loop that learns and a loop that recycles since 1960.
What it does, and how
A loop is treated as a controller and a measurement as its sensor, kept strictly apart so either can be swapped and compared. The repo ships eight loop patterns, each traced to its source in psychology, with runnable controllers, naive baseline sensors, simulated worlds to test them in and a benchmark that scores any sensor you bring.
Where it is up to
Public under Apache-2.0; the demo runs in about five minutes with no API key. It is a teaching playground, not an instrument. Every experiment so far uses simulated data, most of its own qualification gates failed and are published as failures, and it has not yet been run on real agent traces. The validated sensors stay out of the repository.
Where you come in
Clone it, run the demo, and bring a sensor of your own to the scoreboard. If you have real agent traces you could share for a test, that is the missing piece.