The model is not the hard part[Siemens]

The useful news in automation and AI this month is not a new chatbot. It is where the data is allowed to become understandable.

On 15 October Siemens wrote up a partnership with HighByte. HighByte runs on Siemens Industrial Edge, inside the plant. Siemens says it does not store the industrial data. It processes it, and joins the raw signals with the records that give them a meaning: the order, the batch, the maintenance record. Siemens Intelligence Center X is the layer that then runs governed agents on that modeled data. Their line is that connectivity has largely been solved, and context is what still stops a pilot from becoming production.

The example they publish is Vivix, a flat-glass plant in South America. Siemens says a furnace they value at $120 million went from an expected 18 years of life to 20, using statistical process control and predictive maintenance, with the result shown to operators in plain language. The same note says Vivix has 10 operational agents in production, across manufacturing, quality and maintenance, and plans 50 by the end of the year. The 50 is a plan. The 10 is what they say is running.

• In production 10
• Planned 50

A second Siemens move points at the same door from the SCADA side. The WinCC V8.2 release notes put the release in September 2026 and add an MCP server based on the REST API. They list access to runtime tag values, the archive and alarms, and read access to the engineering configuration. The notes describe that as a way to analyse the plant and the project. They do not describe writing a setpoint.

I would not copy the agent count. I would copy the order of work. Name the tag, the unit, the batch and the state before anyone asks a model what the plant is doing. An agent that cannot say which batch a number belongs to is a fluent way to be wrong.