Frontier Technologies in Industrial Operations 2025

AI agents function in a continuous observe, plan and act cycle Execute by leveraging internal or external tools/systemsAct Evaluate possible actions to prioritize them through reasoningPlanCollect and process data from environmentObserve AgentThe basics of AI agents BOX 4 Source: Boston Consulting Group (BCG).AI agents amplify the impact of large language models (LLMs) by giving them access to tools and enhancing their ability to observe, plan and execute actions.3 Traditional AI algorithms, such as machine learning, are task-specific and require human input for defining tasks, providing data and interpreting results. In contrast, AI agents, once trained, can operate and achieve specific objectives autonomously, continuously observing their environment, planning actions and harnessing tools to execute complex tasks. AI agents function in a continuous observe, plan and act cycle, which makes them particularly valuable for operations. Each step is enabled by interfaces or modules:4 –Observe: Agents collect and process data from the environment, including multimodal data, user input or data from other agents. For example, an agent can perceive deviations in production quality and underlying parameters in real time. –Agent-centric interfaces: Agents require protocols, application programming interfaces (APIs) and specifically designed interfaces to input multimodal data or perceive real-time data from multiple sources. –Memory module: Agents have short- and long-term memory, which allows them to remember general knowledge, past actions and decision-making. –Plan: Agents and their underlying LLMs evaluate possible actions to prioritize them through logical reasoning, in accordance with their objectives. In the example above, the agent reviews possible actions to improve quality and decides to change production parameters. –Profile module: Agents have defined attributes, identities, roles or behavioural patterns. The roles can be predefined, or agents can be flexible and dynamically adapt to new roles. –Reasoning module: Agents have limited reasoning capabilities. The underlying LLM is capable of decomposing the agent’s prompts and returning an actionable plan. It extracts key insights and makes logical connections by replicating reasoning steps observed in training data. This enables agents to decide on the required next steps by breaking down complex tasks into small actions to achieve their objectives. Recent studies have shown that current LLMs are not yet capable of formal reasoning. Real- world solutions thus require other types of AI and solvers and cannot solely rely on existing LLMs.5 –Act: Agents execute actions by harnessing internal or external tools and systems. For example, an agent accesses the machine controller and changes the defined machine parameters. –Action module: Agents decide which tools to use, using access mechanisms such as APIs, system integrations or other agents as needed. Functioning in this cycle, agents continuously learn from self-reflection or external feedback. Through goal-oriented learning approaches, such as reinforcement learning, agents continuously adapt and refine their strategies over time. This makes them particularly valuable in complex, dynamic environments where conditions and objectives are constantly shifting. Such environments can be found widely across industrial operations. As part of multi-agent systems, in which specialized agents work together by dividing complex problems among themselves, they can automate entire processes end-to-end. Frontier Technologies in Industrial Operations 11