Part 5 of 7 in our Edge to Enterprise in practice series. In i3X vs MQTT an AI agent pointed cold at CATIM Nimbus answered plain-English questions about live PLC values with no integration code; the question leaders then ask is the right one: what else may it do?
The question on every leadership agenda
Leadership teams are being asked to let AI agents near operations. The benefits are easy to picture: an agent that answers a supervisor's questions about the shift, spots a drift worth a look, drafts the handover. The risks are just as easy to picture: an agent that changes something it should not, with nobody knowing who asked.
Good governance does not mean choosing between the two. It means answering three questions plainly: who is this agent, what may it do, and what record does it leave?
Reading: useful from day one
An agent connected to the i3X API discovers the plant by itself: what exists, what each value means, its unit, its quality and when it was true. It needs no briefing and no integration code written for the plant. For most of what agents are good at - answering questions, summarising a shift, flagging a trend - reading is all they need, and reading is the safest thing they can do.
Every agent has its own identity and role
In CATIM Nimbus, every client of the i3X API - a dashboard, a script, an engineer's tool, an AI agent - has its own token. Nothing shares a login.
- The role decides what it may do. A viewer reads and subscribes. An operator may also change published setpoints and send everyday commands. An engineer may also send the few commands that discard data.
- Access ends when you say so. Removing a client's token ends its access at once, open data streams included, with no restart.
- One agent cannot crowd out the rest. Each client's subscriptions belong to it alone and are capped, so a runaway agent cannot swamp the unit.
- It fails closed. If the unit cannot read its list of clients, it lets no one in.
Refused, and written down
An agent with a read-only role that tries to write is refused, and the refusal is logged: which client, which value, what it tried. That record matters as much as the refusal. An agent that keeps attempting what it may not do is something you want to know about, early.
When an agent may act, it acts within limits
Some agents need to act: adjust a setpoint, apply a prepared job. CATIM Nimbus allows that only through narrow, checked paths:
- Only what is published as writable. A setpoint is writable only if it was published that way, with its range and the least role that may change it. A value outside its range is refused, and a write is confirmed only once it has been read back.
- Commands from a short, fixed list. Apply a job, load one, change the mode, set the logging level: each command has the least role that may send it, and its argument is checked before it reaches the controller, which checks it again.
- No log, no command. The intent is logged before a command is sent, and the outcome when it returns.
- Nothing else is writable from outside. The control program's own variables are not on the list.
And a setpoint written by an agent still reaches a plant output only where a Tributary flow is armed to use it and the controller's own enable allows it. The agent's role decides what it may ask; the controller decides what happens.
Rehearse on a simulated machine
Before an agent meets a real line, its plan can be rehearsed against a simulated machine served through the same i3X API. That is how we test the platform ourselves: on a simulated machine, not a customer plant, an agent set a fault and read back, through the controller, what the controller made of it, with nobody at the switches. The same client code reads the simulated device and the real controller, so the rehearsal uses the very interface the agent meets on the line.
What it means for you
- For executives: a governance story you can put in front of your board: each agent named, each role limited, every attempt recorded.
- For project managers: start every agent as a viewer, and widen a role only when a use case needs it, one agent at a time.
- For engineers: no agent writes into your control program. What is writable is what you published as writable, with its range.
Bring us your first use case
If you are being asked to let AI agents near your plant and want an answer you can defend, we would like to help you frame it. Get in touch.
About CATIM. CATIM - Creating Automation Technology In Motion - empowers people to engineer automation solutions that solve hard technical problems. CATIM Nimbus, the CODESYS edge controller with advisory AI, is the next generation of the Nimbus platform it has built since 2019.