Agents that do the routine work, inside limits you write down.
Automation near an operating plant fails one of two ways: it does too little to be worth having, or it does something nobody authorised. So the work gets graded by consequence. Reading, checking and drafting run on their own. Moving a set point, closing a permit or touching the safety system stops for a person.
What an agent may do, and where it must stop.
Each level is written down, agreed before deployment, and attached to the agent rather than to a policy document nobody reads. Moving an agent up a level is a decision your team makes deliberately.
- Observe
- Reads the record and reports what it sees. Changes nothing.
- Recommend
- Proposes an action with the evidence behind it. A person decides whether to take it.
- Draft and route
- Prepares the work (a work order, a scope item, a permit request) and sends it to a named approver.
- Act with approval
- Executes only after the named person signs, and records who signed.
- Never
- Anything irreversible, anything that moves a set point, anything on the safety instrumented system, anything outside the written limits.
The agent runs to the gate. It does not run through it.
Reading, checking and drafting happen on their own. The step with consequence waits for a person, by design rather than by configuration.
A trigger fires: a threshold, a schedule, an event. The agent pulls the readings, runs the checks its charter allows and drafts the work, attaching the evidence it used. All of that is reversible, so none of it needs a signature.
Then it stops. The draft sits with a named approver, and the actuator on the other side of that gate is not something the agent can reach. What happens after the signature is recorded against the person who gave it.
Every run can be replayed.
Each step an agent takes is recorded: what it read, what it concluded, what it did. A run can be replayed after the fact, and the changes it made can be undone. There is a visible stop, and limits on how much an agent can do before it has to pause and ask.
Your questions, answered
- What stops an agent doing something it should not?
- Its written limits. An agent reaches only the tags and the tools its charter names, and anything irreversible stops for a named person no matter how confident it is. The safety instrumented system is not on that list at any level.
- Can we see what an agent did?
- Yes. Every run is recorded step by step and can be replayed, and its changes can be rolled back. An action nobody can audit is an action nobody should have authorised.
- Is there a stop?
- Yes, and it is visible rather than buried. Agents can be paused or halted mid-run, and they have limits on how much work they can do before pausing for a person.
- Who decides what an agent is allowed to do?
- You do, before it runs. The levels are agreed at deployment and each one names the person who approves at that level.