Agentic AI Framework for Manufacturing Operations: Gilad Langer - Head of Digital Manufacturing Practice, Tulip Interfaces
The promise of AI agents in manufacturing is about creating systems that can actually adapt when your supply chain gets disrupted, when a machine fails, or when customer demand shifts overnight. But here's the problem: without a clear framework, you end up with AI pilots across different parts of the plant, each solving local problems, none of them working together. A collection of disconnected bots, overlapping efforts, and a governance nightmare.
โ
Gilad Langer, Head of Digital Manufacturing Practice at Tulip Interfaces, has spent 30 years, starting with his PhD research in the 1990s on multi-agent systems, working on this exact problem. His recent framework for Composable Agentic AI in Manufacturing Operations offers a fundamentally different approach to data architecture and governance. More importantly, it provides a practical path forward for organizations trapped between their legacy systems and the promise of AI-driven operations.
โ
Why Manufacturing Needs An Agentic AI Framework
Manufacturing operations are what systems scientists call "complex adaptive systems", they share more in common with traffic patterns and weather systems than they do with customer service chatbots. These systems are inherently chaotic, but not in a bad way. They have patterns, and those patterns can be influenced.
Think about the Toyota Production System. Toyota figured out decades ago that manufacturing behaves like a complex system. Their insight? Don't try to control everything from the top down. Instead, create simple rules that prevent the system from spiraling into bad patterns. Pull instead of push to reduce bottlenecks. Remove obstacles immediately through on-demand problem solving. Create flow rather than fighting against the natural dynamics of the system.
This matters because AI agents work the same way. Each agent is a discrete entity following its own goals, working autonomously but interacting with others. When you put multiple agents together, you get another complex adaptive system. And here's where it gets interesting: if you use a complex adaptive system (your AI agents) to manage a complex adaptive system (your manufacturing operations), you can get the best of both worldsโadaptability plus control.
But only if you have the right framework.
โ
A Data Architecture for AI Agents in Manufacturing
Before you can deploy agents effectively, you need to solve a fundamental data problem. Traditional manufacturing data models are too complicated. They try to capture everything, the physical objects, the transactions, the relationships, the history, all in rigid database structures that require a data scientist to interpret.
The Artifact Model takes a different approach. Walk into any manufacturing facility and ask: what do we actually have here? You'll get a surprisingly short list:
Physical artifacts: machines, tools, rooms, areas, materials, work-in-progress, finished products. Things you can touch.
Operational artifacts: orders, defects, tasks, events, schedules. Things you do with or to the physical stuff.
That's it. Every manufacturing plant, regardless of industry, operates with roughly 10-12 types of artifacts. A CNC machine and a testing device? They're 80% the same from a data perspective. Different specific attributes, sure, but the core structure is identical.
When your operators, engineers, and agents can all look at the same data structure and immediately understand what they're seeing, you've solved the democratization problem. No more waiting weeks for someone to write a custom query or generate a report. The complexity of your data model should never exceed the complexity of what you're actually making.
This means your agents have a shared vocabulary. A machine agent knows how to find its maintenance history. A product agent can query its quality parameters. A schedule agent understands which resources are available. They're all working from the same playbook.
That's it. Every manufacturing operation, regardless of complexity, boils down to these categories. Most facilities have fewer than 10 distinct physical artifact types and a similar number of operational artifact types.
Here's why this matters:
Simplicity enables democratization. When your data model reflects the actual shop floor rather than abstract database optimization, engineers and operators can understand it. They can build agents. They can govern data quality. You're not the bottleneck anymore.
Templates enable scale. Yes, a CNC machine and a test stand are different. But 80% of their attributes are identicalโlocation, status, maintenance history, performance metrics. You create a common template for "machines" with specific extensions. Your artifact model grows organically but stays manageable.
Relationships become intuitive. Instead of complex foreign key relationships, you have natural connectionsโthis material is processed by this machine, this task is part of this order. Knowledge graphs build themselves. AI agents understand context without complex joins.
History separates from structure. The artifact model defines what things are, not what happened to them. All the transactional dataโyour UNS streams, your historian data, your event logsโlinks to artifacts by ID. Agents can pull their entire history when needed without bloating the core model.
This is fundamentally different from trying to make traditional MES or ERP data models work with AI. Those systems were designed when data storage was expensive and computing power was limited. The artifact model assumes modern capabilitiesโcheap storage, fast queries, and AI that can make sense of unstructured history.
โ
Types of AI Agents in Manufacturing Operations
With your data foundation in place, you can deploy agents strategically. The framework identifies four categories, each serving a specific purpose:
Physical Agents represent actual objects on your shop floor:
Machine agents monitor equipment health, track performance metrics like OEE, and predict failures before they happen
Product agents follow individual units through production, maintaining quality data and genealogy
Tote agents track material movement, making it trivial to find components and maintain traceability
Operational Agents manage workflow and respond to events:
Order agents oversee entire production orders from start to finish, tracking progress and material consumption
Deviation agents activate when something goes wrong, classifying issues and triggering appropriate responses
Schedule agents dynamically adjust production plans based on real-time conditions
System Agents handle integration with your existing infrastructure:
ERP agents manage the data flow between your production platform and enterprise systems
UNS agents enable real-time data exchange across your entire operational landscape
Data lake agents ensure production data flows to your analytics systems for model training and insights
Device agents connect sensors, scanners, and instruments seamlessly
Staff Agents augment human capabilities:
Quality research agents help operators find documentation and troubleshooting steps instantly
App builder agents generate templates and suggest structures, accelerating development for citizen developers
The key insight: these agent types align with your Artifact Model. A machine agent isn't trying to understand everything about your plantโit's focused on one machine, represented consistently in your data layer. This bounded scope is what makes agents practical and safe.
โ
Practical Implementation of Agentic AI in Manufacturingโ
The biggest challenge isn't technical, it's cultural and organizational. Manufacturing leaders face a paradox: how do you govern a system designed for bottom-up emergence without crushing the adaptability that makes it valuable?
The answer comes from understanding composability. True composable systems have five characteristics: bottom-up development, iterative improvement, lean operations, democratized creation, and human-centric design. Your governance framework needs to enable these characteristics, not fight them.
Start Absurdly Small
Don't create a plant-wide governance framework before you've deployed a single agent. Pick one critical machine that causes frequent disruptions. Put sensors on it. Create an artifact record. Build one agent that helps operators understand when it's about to fail. This takes hours, not months.
Learn what governance you actually need from this first implementation. Maybe it's "who can create agents for critical equipment?" Maybe it's "what data sources can agents access?" You don't know until you do it.
Build Governance Capabilities as Patterns Emerge
After three or four agent deployments, you'll see patterns:
Certain types of agents are universally useful (create templates)
Some data sources need access controls (implement security)
Agent interactions need logging (add observability)
Some agents need human approval before action (build workflow)
Each governance capability solves a real problem you've experienced, not a theoretical concern. This keeps governance lean and relevant.
Focus on Agent Quality, Not Agent Count
Traditional metrics ask "how many systems have we deployed?" Agentic systems need different measures:
How quickly can operators get answers from agents?
Do agents have access to the data they need?
Are agent recommendations being followed or ignored?
When agents fail, how fast do we detect and respond?
You're governing a living system, not managing a project portfolio.
Embrace the Timeline: Hours for Impact, Months for Scale
If someone asks how long it takes to connect a machine and deploy an agent that delivers value, the answer should be "hours." If they ask how long to transform plant-wide operations, the answer is "many months of iteration."
This is the opposite of traditional implementations where you spend months in design before seeing any value. The time investment shifts from up-front planning to continuous improvement.
Mandate Platform Composability
Here's the hard truth: you cannot do this with traditional MES, QMS, or ERP systems. Those platforms were built for the opposite philosophyโcentralized control, up-front design, change management. Trying to retrofit them for agentic AI is like trying to convert a mainframe into a cloud-native microservices platform.
Use Denga's test: Ask if your platform supports bottom-up development, truly democratizes content creation, enables lean iteration, and maintains human agency. If the answer to any of these is "well, with some customization..." you're fighting the wrong battle.
The platform question isn't about vendor preferenceโit's about architectural compatibility. Your platform needs to be designed for agent-based operations from the ground up.
โ
Conclusion
The shift to agentic AI in manufacturing isn't primarily a technology challenge, it's a data architecture and governance challenge. The hard questions aren't about which AI models to use, but about:
How do we structure data for agent autonomy while maintaining system coherence?
How do we govern bottom-up creation without losing control?
How do we measure system health when behavior is emergent rather than designed?
How do we train organizations to think in agents rather than applications?
These are exactly the kinds of strategic questions data and analytics leaders need to answer. The frameworks exist. The technology is ready. The question is whether your data architecture can support it.
Start with one machine, one agent, one use case. Learn what your organization actually needs rather than what you think it needs. Build your governance framework from real experience, not theoretical concerns. And most importantly, accept that the goal isn't to design the perfect system up-front, it's to create a system that gets better every day through emergence and adaptation.
That's how nature builds systems that survive and thrive through constant change. That's how manufacturing needs to work in an unpredictable world. And that's the opportunity for data leaders who are willing to rethink their fundamental assumptions about architecture and governance.
โ
ย
