Agent products differ less in capability than in autonomy. A copilot proposes and a human executes. An autonomous agent executes and, at best, reports. The same underlying model can be deployed either way, and the choice is an operational risk decision rather than a technical one.
The copilot pattern
The agent drafts the analysis, the query, the report section or the recommended action. A person reviews and commits it. Throughput improves, and the accountability structure is unchanged: a named human still made the decision.
This matters more in water than in most sectors, because decisions are frequently defended later. Regulators, auditors and occasionally courts ask who decided and on what basis. "The system decided" is not an answer an operating licence survives comfortably.
The autonomous pattern
Autonomy is justified where the action is reversible, the volume is beyond human review, and the cost of an error is bounded. Some data work qualifies: flagging suspect sensor readings, gap-filling a record with the fill marked as inferred, generating a first pass anomaly list overnight.
Very little operational actuation qualifies. Not because agents are unreliable in principle, but because the failure is correlated: an agent with a wrong premise makes the same wrong decision consistently, quickly, across the network, which is a different risk profile from a human making occasional independent errors.
A practical ladder
The useful framing is a sequence rather than a binary.
- Read only. The agent answers questions, touches nothing. Almost always the right starting point.
- Draft. It produces something a person signs: a report, a work order, a query.
- Act with confirmation. It proposes a specific action and executes on approval.
- Act within a bounded envelope. It acts alone inside explicit limits, with every action logged and reversible.
Most utility value sits in the first two rungs, and most failed deployments started at the fourth.
What to insist on in either case
An audit trail of what the agent saw and did. A stated scope of what it may touch. A way to reproduce an answer later. If a vendor cannot show those three, autonomy is not the thing to be arguing about yet.
We take the same line on agents against dashboards: the interface is the easy part, and the parts that make it trustworthy are where the work is.