Every water network already runs on automation. Pumps start on level setpoints, valves modulate on pressure, alarms fire on thresholds. This is rule-based control, it is decades old, and it is extremely good at what it does.
The question is not whether to replace it. It is which decisions, if any, should move to an LLM agent.
What rules do that agents cannot
A rule is deterministic, fast, verifiable and cheap. Given the same inputs it produces the same output every time, and an engineer can read it and say whether it is correct. For anything touching plant safety or actuation, those properties are not negotiable, and no current agent architecture offers them.
A rule also fails predictably. When a setpoint is wrong, it is wrong the same way every time, which makes it findable. An agent that is wrong occasionally, for reasons tied to phrasing or context, is far harder to debug.
Where rules run out
Rules handle conditions someone anticipated. They degrade in three situations:
- Combinatorial conditions. A rule per fault is tractable. A rule per combination of faults is not, and the interesting failures are combinations.
- Alarm storms. During an event, a rule set produces hundreds of true alarms with no ranking. A human then does triage under time pressure, which is exactly when triage is worst.
- Unstructured context. Maintenance history, operator notes, a contractor email about a closed valve. Rules cannot read these; they are precisely what an operator uses to interpret an alarm.
The split that works
Keep actuation deterministic. Put the agent where a human currently reads and interprets, not where a relay currently closes.
In practice: rules detect and act; the agent explains, correlates and proposes. An agent that says "these forty alarms are one burst near node 412, here is the maintenance record and the last three pressure traces" is doing work no rule set does, and it is not touching the network to do it.
The decision rule
If a wrong answer actuates equipment, it belongs in a rule. If a wrong answer wastes an engineer's ten minutes, it can belong to an agent. The boundary is the consequence of being wrong, not the difficulty of the question.
This is the same argument we make about physics models against machine learning: the statistical component earns a place where being approximately right quickly is useful, and not where being exactly right matters.