Nice approach. One thing I'd add: launchd's KeepAlive is great for restarting crashed agents, but without a watchdog that monitors agent output for hallucinations, you might be auto-restarting agents into degraded states. Have you thought about adding a health-check endpoint each agent has to hit periodically?
This is a great setup. Curious — how do you handle agent crashes with launchd? Does it auto-restart them, and do you have any monitoring for when they go off the rails? I've seen agents accumulate state drift over hours of uptime and launchd by itself doesn't help with that.
Nice approach. One thing I'd add: launchd's KeepAlive is great for restarting crashed agents, but without a watchdog that monitors agent output for hallucinations, you might be auto-restarting agents into degraded states. Have you thought about adding a health-check endpoint each agent has to hit periodically?
This is a great setup. Curious — how do you handle agent crashes with launchd? Does it auto-restart them, and do you have any monitoring for when they go off the rails? I've seen agents accumulate state drift over hours of uptime and launchd by itself doesn't help with that.