Rockwell gives AI agents only the simulation tasks a validated digital twin can support
At Automation Fair 2026, Rockwell Automation laid out rules: AI agents may chain digital twin simulation tools only for tasks a validated model supports. Model validation comes first. Green, yellow and red zones set where virtual tests replace physical ones, starting with one or two engineering questions.
This story was produced through MarketScale. See how Industrial IoT teams put it to work with AI Visibility (GEO).
Key takeaways
An AI agent's scope in Rockwell's method is set by how well the digital twin has been validated: green-zone tasks can replace physical tests, yellow-zone tasks can trim some, red-zone tasks are off limits.
People who use simulation results outnumber the engineers who run simulations by about 10 to 1, so agents are aimed at the queue of handoffs between specialists, not at engineering judgment.
A model can match conveyor dimensions and machine motion and still assume perfect part supply, so any green label should be tied to the specific question it was validated for.
Free workspace
Turn your Industrial IoT expertise into content.
Record interviews, organize footage, and write with AI on a free trial of the MarketScale platform for qualifying companies. No demo required, no credit card.
On a typical factory project, about ten people are waiting on every engineer who actually runs a simulation. Rockwell Automation put that ratio in its presentation materials at Automation Fair 2026 in Boston on Nov. 17. The problem it describes is familiar to anyone who has watched a design question sit in a specialist's inbox. Rockwell's proposed fix, AI agents, comes with more conditions than most agent pitches do.
In a session titled "Building Trustworthy Agents for Digital Twin Engineering," Chelsea Barnes, a digital consulting director, and Franklin Lufkin II, a principal solution architect, set out when Rockwell will let agents do engineering work, according to BigGo Finance's account of the session. Agents chain simulation tools into one continuous run. They get only the tasks the underlying model has been validated to support. And customers should start with one or two high-value questions long before they try to cover a whole plant.
The most useful part for a manufacturing engineering leader is where Rockwell draws the line. What an agent may do depends on how well the digital twin has been proven against physical tests. That makes a validated model the first deliverable of any agent pilot. The agent comes after.
One bottle question, two simulation tools
The presenters described a fragmented workflow. A packaging engineer asks a question, and structural analysts, fluid analysts, equipment engineers and controls engineers each take a turn with their own tools. Every handoff adds waiting time, and each tool's output tends to stay disconnected from the others. With users outnumbering simulation specialists about 10 to 1, the queue is built into the org chart.
The demo turned one question into a chain of tasks: how should wall thickness and conveyor speed be set to hit target output without bottles tipping over? The agent first ran finite element analysis to model how the bottle deforms and to find its center of gravity. Then it passed that work to Emulate3D, Rockwell's digital twin software, to test whether bottles topple on the conveyor, and came back with a wall thickness and a conveyor speed. It logged the inputs and results of every step.
That log matters more than it looks. A recommendation an engineer can trace back through each tool is one a reviewer can check. An answer without that trail just looks confident. Barnes' point, as reported, was that agents make it easier to use several tools at once, but only a trustworthy model produces answers worth acting on.
A second demo trained an inspection system before the product existed. The 3D design data for a new bottle went into Emulate3D to generate synthetic images. The agent moved those images into a repository to train a vision inspection system, and the results were fed back into the virtual production line. For launch teams, this could move vision-system training ahead of first physical samples. The quality of that head start depends on how closely the synthetic images match what the real camera will see.
Three zones and a job description
Lufkin argued that validating the model has to come before any agent output reaches the factory floor. In his framing, validation means confirming the model was built correctly: real equipment motion, conveyor dimensions and other physical details are represented as they actually are. Once that work is done, each task falls into one of three tiers.
- Green zone: virtual tests can replace physical tests.
- Yellow zone: virtual tests can reduce some physical tests, not all.
- Red zone: outside what the model supports; virtual results are not reliable.
In Rockwell's method, the agent can only go as far as the model has been proven: green-zone work can stand in for a physical test, yellow-zone work can trim a few, and red-zone work gets no vote at all. The presenters also said agents should get only validated tasks, such as running models or converting and summarizing results. Each agent needs what they called a job description, which limits the data and input ranges it can use.
The rule for changes is just as strict. When a physical line or product changes, the digital model gets updated and revalidated first, and only then does anyone touch the floor. It borrows continuous integration and continuous deployment from software. For plants where a twin was built for commissioning and then left alone, that rule could mean an ongoing upkeep job that nobody owns today. Ownership is worth settling before an agent depends on the model.
Accurate geometry, optimistic factory
A month before Automation Fair, Konstantin Saifoulline, founder and CEO of Airsup, made a related argument in Automation World. He wrote that the most misleading simulation result is often the cleanest. A virtual robot moves fast and output climbs, but the scenario quietly assumes ideal part delivery, no rework, a constant product mix and an operator who is always available. His advice: a model should first reproduce how the existing process actually behaves, closely enough for the decision at hand.
In Rockwell's method, the agent can only go as far as the model has been proven: green-zone work can stand in for a physical test, yellow-zone work can trim a few, and red-zone work gets no vote at all.
This tests Rockwell's framework more than it contradicts it. Lufkin's validation checklist covers physical fidelity: motion and dimensions. Saifoulline's warning is about operating assumptions such as cycle-time variation, stoppages and shared operators. A twin can pass the first check and still fail the second. That suggests a green label should belong to a model and a specific question together, never to the model in general. Saifoulline cites NIST's digital twin research, which treats verification, validation and uncertainty quantification as required for trustworthy decision support, and which judges a model's outputs against their intended use rather than accepting them because the model runs.
The two sources also agree on who decides. Saifoulline writes that an agent may prepare options, flag contradictory inputs and keep a register separating measured values from estimates. It must not release a robot program, change a safety-rated procedure or approve a purchase. Rockwell's version is a similar split: the agent handles repetitive analysis and the connections between tools inside validated limits, and people review the results and their limits.
Plant-scale simulation, question-scale pilots
The day before the session, Rockwell expanded its partnership with Nvidia. It will build Nvidia's Omniverse simulation platform, RTX rendering and OpenUSD into Emulate3D, taking simulation from single machines up to the whole plant. Matheus Bulho, Rockwell's senior vice president of software and control, said the combination lets customers simulate larger and more complex systems. He added that it gives AI systems a foundation to learn, validate decisions and optimize performance before real-world deployment.
There is a tension here, and Rockwell's own advice settles it. A plant-level twin has far more to validate than a single conveyor. The sequence the presenters recommended starts small: validate the model against physical tests, extend it to new production lines, then gradually replace physical testing in the areas where trust is established. Laura Green, senior manager of advanced manufacturing simulation at GE Appliances, described the benefit in practical terms. Exploring ideas virtually lets her team evaluate them earlier and make engineering decisions with enough information behind them.
What to ask before the first pilot
Teams scoping a first agent project with Rockwell or an integrator can draw several pointed questions straight from the session. Which zone does each proposed agent task fall into, and what physical test put it there? What data and input ranges does the agent's job description allow, and what happens when a request falls outside them? When the line changes, who updates and revalidates the model, and how long does that take?
Ask to see the step-by-step log from a pilot run, then check one assumption in it against a week of real line data: part arrival, stoppages, operator availability. If the model only works with perfect supply, it isn't green yet.
This approach matters most where the engineers who interpret results far outnumber the ones who run analyses, and where the question is narrow enough to check against a physical trial. It matters least for a plant that wants an agent to take over line decisions next quarter. Rockwell told customers not to try that. The real test is whether early adopters can name one physical test their green-zone model has actually retired.
Sources
Featured companies
Your experts belong here
Every story in MarketScale Industrial IoT starts with a company putting its controls engineers, plant-floor specialists, and integration partners on the record. Buyers are already reading this topic. The only question is whose experts they find.
Plant and controls buyers research deep before contact, and your engineers get to shape that research.
About the author
The MarketScale Newsroom reports on the companies, technologies, and trends shaping 16 B2B industries. It turns primary sources and expert commentary into clear, useful coverage for the people doing the work.