Research on agentic AI

Agentic AI refers to artificial intelligence systems that operate with a certain degree of autonomy, make decisions, and pursue goals without constant human supervision. Agentic AI systems are based on foundation models with reasoning and acting capabilities and model-driven specialized AI agents that are combined to form multi-agent systems.

Agentic AI can be integrated into complex environments, where it can replace or facilitate existing workflows. In short, agentic AI creates added value by enabling companies to let AI independently take on real-world tasks, making processes more efficient and responsive.

Deploying autonomous AI agents in professional environments requires a high degree of reliability, transparency, and operational control. At Fraunhofer IAIS, our research focuses on bridging the gap between rapidly evolving scientific breakthroughs and the stringent demands of industrial applications. We design agentic systems that integrate seamlessly into complex enterprise architectures while strictly complying with data sovereignty, security, and regulatory standards. 

Our research addresses the following core challenges:

• Specialized reasoning: How can domain‑specific large language models be engineered to support structured reasoning, formal decision‑making, and knowledge‑intensive analysis in fields such as law, medicine, and engineering? We develop methods to enhance domain fidelity, logical coherence, and factual robustness so that agentic systems can handle complex, multi‑step tasks with verifiable accuracy. 
• Edge & sovereignty: How can we design agent systems that run on‑premise, at the edge, or in sovereign cloud environments to ensure privacy, reduce latency, and control inference costs? Our research explores decentralized architectures and secure execution environments that maintain full control of sensitive data while ensuring high‑performance agent behavior.
• Reliable evaluation: How can we benchmark the performance, safety, and robustness of autonomous agents in dynamic, real‑world environments? We develop evaluation frameworks aligned with European standards and industrial requirements, enabling systematic assessment of long‑horizon reasoning, tool‑use patterns, risk profiles, and emergent behavior. 
• Enterprise MLOps: What are the optimal architectural frameworks for deploying, monitoring, and maintaining agentic systems in production? We design agent‑centric MLOps pipelines that ensure reproducibility, observability, and governance across the entire lifecycle. This includes data‑quality optimization for agent‑driven workflows and mechanisms for continuous adaptation and compliance. 
• Extended capabilities: How can agentic AI be expanded beyond classical task automation by integrating simulation environments, operations research, and physical AI systems? We investigate hybrid setups that enable scenario planning, real‑time optimization, and embodied intelligence, thus broadening the application space of agentic systems across industrial sectors.