Managing agent loops is all the rage. I'm building an app that needs extensive data from all colleges in the US and ingests from a standardized source, CDS.
1. One agent goal runs the CDS ingestion
2. One agent goal improves CDS speed toward 4x
3. A supervisor goal watches both, restarts them if they stall, and edits their prompts if the instructions are what caused the stall.
Lastly, an automation wakes the supervisor every ~20 min to check on the worker agents.
Basically: goals do the work, supervisor keeps them alive, automation keeps the supervisor awake.