Adaptive, Agent-Oriented Control for Biomanufacturing Systems

Agentic AI goes beyond predictive and generative AI and, in biomanufacturing, has the potential to enhance efficiency by integrating with existing manufacturing infrastructure such as IoT sensors, process information management systems, management execution systems, and even enterprise resource planning software. The challenge, however, is that industrial biomanufacturing processes are complex, demand resilience, and are tightly regulated.

The Adaptive Agent-Oriented System Control (AAOSC) framework developed by a team from the Technical University of Denmark (DTU) and SiC Systems addresses that challenge through a decentralized control layer. In it, “specialized autonomous agent ‘hives’ [are] coordinating digital twin enabled manufacturing infrastructure and real-time communications protocols.” The latter lets biomanufacturers integrate models, make learning-based inferences, and control process systems.

Four AAOSC case studies were discussed in a recent paper by Seyed Soheil Mansouri, PhD, professor at DTU and co-founder and CSO of SiC Systems and Christopher J. Savoie, PhD, co-founder and CEO of SiC Systems, and inventor of the agentic AI technology behind Siri. Those case studies “demonstrate AAOSO’s prowess [in] reducing deviating durations, averting shutdowns in severe fault scenarios, and boosting efficiency through virtual quantum and classical sensing and decentralized reasoning, all while aligning with regulatory imperatives…”

Despite its capabilities in monitoring process, identifying discrepancies, and recommending solutions, agentic AI “is not yet fully ready for complete, independent control in biopharmaceutical manufacturing,” Mansouri tells GEN. “Any AI that directly affects medicine quality still needs strong human oversight and full approval. We are getting closer, but full integration requires official [regulatory] clearance.”

The AAOSC framework that Mansouri and colleagues built may be unique in the industry. It isn’t all-knowing and “God-like,” he points out. Instead, “our methods are grounded in physics, chemistry, and biology within an agent ‘hive’—an orchestration of rule-based, mathematically informed agents. So, AAOSC is, foundationally, a different philosophy of building AI [in which] humans are in control.”

First, run in shadow mode

To introduce agentic AI, Mansouri advises starting gradually. “Run the AI alongside your current control systems in shadow mode—it watches everything and gives recommendations, but doesn’t make any actual changes without human oversight. This lets the teams learn how it works without any risks to production. Once confident, you can slowly expand its role while always keeping humans in final control.”

Both the FDA and EMA require systems that are fixed rather than continuously learning, he points out, and that can complicate adoption. To minimize the potential for regulatory issues that may arise by integrating AI into manufacturing processes, “work closely with your quality and regulatory teams from the beginning.

“Always maintain clear human responsibility, so no one is left wondering who is accountable if something goes wrong. Strong cybersecurity is essential,” Mansouri adds, “because these AI agents connect and talk to each other.” Therefore, “Start small, test thoroughly, and talk to regulators early.”

The post Adaptive, Agent-Oriented Control for Biomanufacturing Systems appeared first on GEN - Genetic Engineering and Biotechnology News.