loader image

Home → Projects →

Agentic Engi­neering in the Soft­ware-Defined Brewery

The Chal­lenge
We Tackled

OT engi­neering still means moving by hand across a chain of discon­nected tools: from P&ID and DEXPI, through PLC and HMI engi­neering, to deploy­ment and asset manage­ment. Every hand-off costs time and invites error. The first wave of AI agents in engi­neering made this worse in one way: node-graph agent pipelines are hard for an engi­neer to follow, and hard to trust. The chal­lenge was to compress the engi­neering life­cycle with agentic automa­tion while keeping a human engi­neer firmly in the loop, able to see exactly what the agents are doing.

Who is
Partic­i­pating?

And who was doing what?

  • Boehringer Ingel­heim: end customer, bringing the process-industry require­ments and oper­a­tional perspec­tive
  • KIT (Karl­sruhe Insti­tute of Tech­nology): system inte­grator, tying the compo­nents into one working stack
  • logic­cloud: Process.AI plat­form and the Margo fleet manager driving deploy­ment
  • COPA-DATA: zenon Engi­neering Studio and the zenon (virtual) PLC running the control logic
  • Endress+Hauser: process instru­men­ta­tion and edge hard­ware; sensors, edge gateway, and testbed gate­keeper
  • FLECS: Margo-compliant edge agent running the deployed work­loads, plus the tool­chain turning specs into deploy­able container apps
  • Margo: the open inter­op­er­ability stan­dard tying the inde­pen­dently built compo­nents together

Why this project
matters and for who

For plant owners and oper­a­tors, it’s simple: speed and cost. An orches­trated agent process shrinks the OT engi­neering life­cycle and cuts the time between “we have an idea” and “it’s running in the plant.”

For lead engi­neers, it’s about staying in control. Instead of babysit­ting an opaque agent pipeline, the engi­neer becomes the director of the agentic team. Process knowl­edge lives in a persis­tent, visual BPMN work­flow rather than being buried in a node graph, so over­sight doesn’t cost extra cogni­tive load.

For the wider process industry, this is proof that a soft­ware-defined, cloud-native automa­tion stack isn’t a slide-deck idea. It runs, today. Because it’s built on Margo, it’s not locked to one vendor’s tools, so the approach can scale across the industry rather than staying a single compa­ny’s party trick.

what we’ve
learned

Agentic engi­neering genuinely shortens the OT life­cycle. From P&ID to edge deploy­ment, orches­trated agents cut out manual effort at every hand-off.

Trans­parency is what makes AI-driven engi­neering trust­worthy, not just fast. Putting process knowl­edge into a persis­tent, visual BPMN arti­fact, instead of an opaque node graph, is what let the engi­neer stay in control rather than just watching.

The Soft­ware-Defined Brewery isn’t a future promise. Margo-compliant devices, zenon, and agentic AI already form a working, end-to-end SDA stack today.

And the cloud-native foun­da­tion has to come first. Skip it, and the more advanced agentic use cases simply stay out of reach.

What we’ve
achieved

We built an end-to-end demon­strator that turns a plain-language process descrip­tion into a produc­tion-ready deploy­ment, with no manual hand-offs between engi­neering tools along the way.

The use case: temper­a­ture moni­toring on Tank A via sensor TT-101, 0 to 30 °C span, alerts at 7.0 and 13.5 °C, protec­tive actions at 5.0 and 15.0 °C. An engi­neer described the process in plain language, and from there an orches­trated team of AI agents took over and:

  • Struc­tured the descrip­tion into an executable BPMN work­flow
  • Gener­ated the instru­ment tag spec­i­fi­ca­tions
  • Config­ured the control logic in the zenon Engi­neering Studio
  • Deployed a container­ized PLC to a Margo-compliant edge gateway via the fleet manager

The result was a real, running control appli­ca­tion, not a mock-up. As we like to put it: we turned a spec­i­fi­ca­tion into results. In this case, beer — Germans, am I right.

Every step in that chain maps to an open inter­face or model, which is exactly what let inde­pen­dently built compo­nents from six different orga­ni­za­tions inter­op­erate end to end: P&ID/DEXPI for the spec­i­fi­ca­tion, BPMN for the process model, MCP for agent orches­tra­tion, the zenon engi­neering tool­chain for config­u­ra­tion, and Margo for fleet-to-edge deploy­ment. Field connec­tivity to the sensor runs over PROFINET via Ethernet-APL.

What’s
next?

The next public showing is SPS 2026 in Nurem­berg this November, the indus­try’s leading automa­tion trade fair. Beyond that, the plan is to move from a non-crit­ical demon­strator toward higher-crit­i­cality process control: following the stag­gered path from NAMUR Open Archi­tec­ture moni­toring toward a defined “CPC1” complexity level for core process control, with a broader agent toolset and a deploy­ment stack hard­ened for regu­lated, CRA-compliant oper­a­tion.

curious To
learn more?

No single vendor could have built this alone. Turning a plain-language spec into a running plant took a work­flow plat­form, an engi­neering studio and PLC, a Margo-compliant fleet and edge layer, and an open stan­dard to hold it all together. That’s exactly the kind of cross-vendor collab­o­ra­tion the Open Industry 4.0 Alliance exists for.

If you’re working on soft­ware-defined automa­tion, agentic engi­neering, or Margo-compliant edge deploy­ment, get in touch. We’d love to compare notes and build the next demon­strator together.

Contact