At MES & Industry 4.0 in Porto, Francisco Almada Lobo, Chief Executive Officer of Critical Manufacturing, explains how manufacturing execution systems are evolving fast. The integration of AI agents, reasoning models and memory-driven intelligence is transforming decision-making on the factory floor.
The traditional MES is no longer enough. Static logic, however extensible, cannot keep pace with the accelerating complexity and volatility of today’s manufacturing environments. Whether managing deviation in semiconductor production, responding to machine anomalies in medical device lines, or optimising high-mix production schedules in electronics, the need for intelligent, adaptive control has become unavoidable.
Francisco Almada Lobo, Chief Executive Officer of Critical Manufacturing, believes the era of deterministic rules and static workflows is coming to an end. “We need to stop thinking about MES as a system of record and start thinking about it as a system of intelligence,” he says. That intelligence will not come from human-authored code alone. It will emerge from AI agents, embedded reasoning, and self-adjusting systems capable of operating at speed and scale far beyond human capacity.
The narrative unfolding here is not simply a case of layering AI on top of legacy systems. Instead, it is a fundamental reimagining of how manufacturing software should behave in a world defined by change.
AI agents need more than data
For decades, the mantra in digital transformation has been that data is the new oil. Yet, as AI capabilities have evolved, that metaphor has begun to collapse. Oil requires refinement, and so does data, but AI needs more than clean input. It requires structured relationships, contextual awareness, and computational infrastructure capable of delivering insights in real time.
“Without clean, connected, contextualised data, none of the shiny stuff works,” Almada Lobo says. This is not merely a technical inconvenience. It is the single most significant constraint on the generation of AI value in manufacturing. Most enterprises, he argues, are still stuck in wave one of AI adoption, using traditional machine learning algorithms trained on rigid datasets that are poorly integrated and time-consuming to manage.
This first wave, defined by regression analysis, clustering and decision trees, promised a future where engineers would uncover hidden patterns, optimise yield, and reduce waste. But the practical reality has been sobering. According to Almada Lobo, data scientists spend 80 per cent of their time on data preparation tasks, extracting, cleaning, and organising, leaving little time for the value-creating work of model training and insight generation.
The shift to modern AI, particularly large language models and AI agents has only compounded the architectural mismatch. These models are exponentially more data-hungry. They require not only vast volumes but also diverse data types: logs, sensor streams, maintenance records, human inputs, and more. Legacy relational databases and siloed MES modules are wholly inadequate for the task.
This is where the so-called ‘holy trinity’ of smart manufacturing emerges: the seamless integration of Manufacturing Execution Systems (MES) for orchestration and context, Internet of Things (IoT) platforms for data acquisition, and modern manufacturing data platforms for scalable ingestion and transformation. Each must be fully interoperable, feeding intelligence into the system while capturing new data generated by AI itself. “You need all three layers working in harmony,” Almada Lobo says. “Otherwise, your AI will be flying blind.”
From language to logic
If wave one was machine learning and wave two was large language models, the industry now stands on the edge of a third and even more consequential transformation. Almada Lobo describes this as the age of AI agents, autonomous systems with embedded reasoning, memory, and goal orientation.
Yet, to appreciate the significance of AI agents, it is worthwhile to examine the limitations of current language models. LLMs have transformed how we engage with information, but their application in manufacturing has been met with scepticism. The reasons are practical, not philosophical.
“They hallucinate, they forget everything between sessions, and they expect the human to do all the hard work,” Almada Lobo says. These flaws, hallucination, memory loss, and lack of integration, are not just nuisances. In a manufacturing environment, they are deal-breakers. An LLM that cannot recall operational context or explain its reasoning cannot be trusted with quality-critical processes or regulatory compliance.
To address these limitations, three enabling techniques have emerged. Retrieval-augmented generation (RAG) allows LLMs to access a curated knowledge base, such as MES documentation or past production events, enriching their responses with relevant context. Fine-tuning offers deeper adaptation but comes at a cost. “Fine-tuning a model like GPT-4 can run into millions of euros,” Almada Lobo notes. Even then, it must be repeated for every new model iteration. Finally, prompt engineering provides a lightweight mechanism for steering model behaviour through creative input design, though it remains an art as much as a science.
None of these solutions alone is sufficient. Together, however, they point to an emerging capability: the ability to use LLMs not as glorified chatbots but as reasoning engines that can support or even replace certain MES functions.
The arrival of AI agents
True transformation arrives with the next step: embedding those models into agents capable of taking autonomous, goal-oriented action. The agentic model represents a fundamental shift in software design. Instead of scripting a sequence of steps for the MES to follow, manufacturers define a goal. The agent then selects its tools, queries the relevant data, evaluates its progress, makes adjustments as needed, and iterates toward success.
This is not hypothetical. It is already happening. Frameworks such as AutoGPT and Google’s ReAct enable developers to build agents with long-term memory, planning capabilities, and the ability to interact with external systems. Recent protocols, including Anthropic’s Model Context Protocol (MCP) and Google’s Agent-to-Agent communication standard, are making it easier for these agents to interact with existing enterprise software.
In manufacturing, this means providing agents with access to MES logic and datasets via API, exposing rules, exceptions, material flows, and performance metrics. “The MES becomes an accessible substrate for intelligent reasoning,” Almada Lobo explains. “We are no longer limited to what has been pre-configured. The system can now evolve on its own.”
This evolution unfolds across three levels. The first is basic automation, applying deterministic rules to well-understood conditions. The second is agentic logic, where AI dynamically adapts rules based on the current status and historical patterns. The third and most advanced stage is a full AI agency, where agents reschedule tasks, optimise resource allocation, and adapt to disruptions in real time without predefined workflows.
Managing multi-agent systems
Scaling these capabilities requires more than smart individual agents. As manufacturing grows more complex, so too must the systems of coordination that support decision-making across multiple agents and tools. Almada Lobo describes the emergence of multi-agent ecosystems, where each agent has a narrowly defined scope, such as material management, equipment health, and exception handling, but collaborates with others as needed.
To enable this, agents must be aware of each other’s capabilities. Google’s Agent-to-Agent protocol addresses this by allowing each agent to register with a shared server, thereby announcing its skills to the rest of the system. Tasks that span domains, such as adjusting a production plan based on incoming material delays, can then be distributed dynamically across agents.
At the highest level sits the meta-agent. This orchestrator monitors overall system behaviour, identifies patterns or conflicts, and delegates responsibilities. It also manages the continuous learning loop: agents generate suggestions, humans review them, outcomes are assessed, and memory is updated. Over time, this cycle becomes self-reinforcing. “The system improves with every iteration,” Almada Lobo says. “The agent becomes not only more accurate but more aligned with how your business really works.”
Guardrails, governance and the human role
The promise of autonomous agents cannot be realised without rigorous governance. Almada Lobo is clear on this point. “You cannot give blind autonomy to systems without safety mechanisms,” he says. “That would be reckless.”
The emerging framework includes three critical components. First, guardrails are explicit policy boundaries that define what agents are allowed to do. These could be safety rules, quality thresholds, or regulatory constraints. Second, human-in-the-loop oversight ensures that any action with a significant consequence, such as changing a yield threshold or overriding a compliance rule, requires human approval. Third, decision traceability allows users to see exactly how an agent arrived at its conclusion, complete with visualisations of reasoning paths and token-level analysis.
Recent advances have made this transparency more accessible. Open-source tools now enable graphical tracing of agent decisions, exposing not only outcomes but also the steps and assumptions behind them. “It is no longer a black box,” he says. “You can inspect, validate, and refine the reasoning.”
Acceleration and the new leadership challenge
The most important takeaway from Almada Lobo’s analysis is the pace of change. In less than three years, the industry has moved from early experimentation with GPT-3 to the practical deployment of autonomous reasoning agents in operational systems. “We are living the Law of Accelerating Returns,” he says, referencing futurist Ray Kurzweil’s prediction that exponential technological growth will compress transformation timelines beyond human comprehension.
For executives, this creates an existential leadership challenge. The gap between what technology can do and what organisations can absorb is widening. Scott Brinker’s Martec’s Law captures this dilemma perfectly: technology evolves exponentially while organisations change logarithmically.
To survive, manufacturers must close that gap. “Becoming more agile is necessary, but it may not be enough,” Almada Lobo warns. “Sometimes you need to be revolutionary. Ask yourself, if you were to rebuild your company from scratch today, what would it look like?”
Final thoughts
The vision of the MES as a thinking system is no longer speculative. It is being prototyped, tested and deployed. The architecture is falling into place. The agents are learning. The data foundations are being laid.
Still, Almada Lobo offers a final provocation. “If you do not believe any of this, that is fine,” he says. “But know this, we are not early. We are already late.”
The industry now stands at a crossroads. Executives who wait for clarity may find that by the time they understand, the window has already closed. Those who act now, experimenting with agents, investing in clean data architecture, and rethinking their logic layers, will be the ones to shape the next era of manufacturing.
As Almada Lobo concludes, quoting W. Edwards Deming: “It is not necessary to change. Because survival is not mandatory.”
