An entry point

Everything in a plant falls back on the same range: optimal state, alarm state, shutdown state. So an alarm is your clearest indication that something needs to be addressed. That part is right. But before agentic work orders were spawning, they took the form of SCADA alarms. And you can skate an alarm team to death. Over time they just learn to isolate what actually requires a reaction down to three to five things. Everything else is status quo, or not big enough to warrant the cost of deploying somebody against the benefit that work order might actually provide. So if the agent just spawns a work order off every alarm, you haven't fixed anything. You moved the fatigue from the screen to the queue. The alarm should be an entry point, not an answer. The event triggers a series of events that goes throughout the system and slowly builds a case. You want to isolate the problem to the wound, not the symptoms. That's the whole issue. And most of that case can already be built from what makes up the behind-the-scenes profile, if the data is granular and organized enough. A given status always breaks down to the laws of physics or the raw materials that make up the facility. High pressure can only be built from where it is currently being supplied. That simple logic branches out through the rest of the facility. The hard part was never the reasoning. It's scoping all the dependencies, and scoping them in a way that subverts the limitations of modern day LLMs. Do that, and you get better at looking back at what you've already collected and saying: this is now more important, or this has dropped in value to the case we're building.

https://ojs.aaai.org/index.php/AAAI/article/view/41453