One of the world’s largest glass container manufacturers faced a persistent challenge: gob formation, the key first step in making properly shaped bottles. With over 90 sensors and 60 control actions to manage in real time, the process had never been automated successfully. Alex, the plant’s most experienced operator, had taken over a decade to master it.

‍All the previous attempts at automation had failed for the same reason. Systems based on rules couldn’t replicate the kind of judgment this process demanded.

‍Partnering with AMESA, the manufacturer took a different approach. They didn’t try to program the system, or to prompt it. Instead, they taught it like any new apprentice.

‍With AMESA and a methodology called machine teaching, they were able to use real data and Alex’s expert knowledge to create a multi-agent AI system that could learn to support real-time decisions with expert-level precision.

‍

Simulating the Process Using Historian Data

‍Just like an apprentice, the system needed practice to become an expert. For AI, the best way to practice is in a simulation based on real data.

‍The team used AMESA to create a custom simulation using data already stored in their OSIsoft PI historian, along with edge case data generated through carefully designed experiments. No new sensors or IoT overhaul were needed, just strategic use of what they already had.

‍The data-driven simulation allowed the AI agent system to practice in a feedback loop, testing out actions and seeing the results, to optimize toward the overall goal – in this case, improving the yield of in-spec gobs.

‍

Teaching Expert Knowledge to AI

With the simulation in place, it was time to build and train AI multi-agent systems using Alex’s expert process knowledge.

‍Just like any expert, Alex had distinct skills and strategies he used to make decisions. These skills and strategies are key to teaching AI, because complex tasks require more than a single ability. For instance, pilots learn takeoff, landing, cruising, and navigating turbulence, not just a single skill called "flying.”

‍Alex broke down the gob formation process the same way. In an interview, he described seven core control strategies that he used to manage the process effectively indifferent situations. These strategies became the building blocks of the AI system.

‍

Building and Training a Multi-Agent System

‍Alex’s strategies were translated into a team of “skill agents,” individual AI modules configured with goals and constraints within the AMESA platform. These were then orchestrated together into agent systems to be trained, tested, and benchmarked.

‍Guided by the expert strategies Alex had defined, each AI system practiced the process millions of times in the cloud, running scenarios in parallel to accelerate learning.The innovation team guided the training, validated outcomes, and worked withAlex to refine agent behavior until the systems beat benchmarks.

‍What took Alex 12 years to master was taught in a matter of weeks.

‍

Deployment on the Plant Floor

From multiple trained agent teams, they selected a few top performers and brought them to the factory. Connected to the edge and integrated into the HMI, these systems didn’t replace operators. They operated alongside them.

‍Like a new hire, the AI system had to pass a series of six tests to be considered “deployment ready." Alex oversaw the entire testing process.

‍On test six, Alex suddenly stopped the trial.

‍“Wait,” he said. “I wouldn’t do that.”

‍He paused, looked closer at the decision the system had made, and said “Wow. I never knew you could do it that way.”

‍A strategy emerged from the agent system that even Alex hadn’t considered. The student had taught the teacher.

‍

The Result

The winning system was deployed onto the factory floor and Alex certified it as an expert operator.

‍Instead of working to replace Alex’s hard-earned expertise, the agent system captured, scaled and operationalized it, becoming a tool for Alex and, more importantly, a tool for less experienced operators to use to maintain high-quality production.

‍In just nine months:

• The system passed six real-world validation trials

• It met certification standards set by plant experts

• It reached production readiness — making expert-level decisions with full transparency

‍It was machine teaching — not black-box magic, but intentional, transparent, human-guided AI design.

‍

Build Your Own Production-Ready Multi-Agent System

This approach isn’t exclusive to Alex’s team. It’s now available to any innovation leader ready to turn deep operational knowledge into intelligent systems.
‍

With AMESA’s structured pilot, you can go from use case to deployment in as little as twelve weeks.

‍Here’s what’s included:

• Two AMESA platform seats

• Support to identify a high-value use case and build a simulation from data

• Design and training of AI agent systems based on a structured expert interview process

• Validation, testing, and deployment into production

Contact us to launch your pilot — no new vendor approval or billing change needed.