Transforming Engineering Systems: The Evolution of AI from Tools to Autonomous Agents

Until very recently, AI in engineering functioned primarily as a highly specialized, responsive tool. An engineer provided a set of instructions or boundary conditions, the system executed the calculation, and it quietly waited for the next command. Generative AI expanded this dynamic by producing complex code, technical documents, and preliminary analysis. However, it remained bound to a fundamental prompt-and-response architecture.

Agentic AI represents a distinct departure from this passive framework. Rather than waiting for step-by-step instructions, agentic systems are designed as autonomous entities capable of pursuing complex objectives with minimal human intervention. These architectures exhibit proactive planning, long-term contextual memory, integration of external tools, and dynamic behavioral adaptation within unpredictable environments (Abou-Ali et al., 2025; Acharya et al., 2025). Instead of merely answering inquiries, an agentic system makes multi-step action sequences, directly interfaces with external sensors, databases, and actuators, and autonomously modifies its strategy based on operational feedback.

The shift from static software tools to goal-driven agents is rapidly reshaping the enter-prise and industrial landscape. Gartner identified agentic AI as its top strategic technology trend for 2025, projecting that by 2028, at least 15% of daily workplace decisions will be made autonomously—up from nearly zero in 2024—and that 33% of enterprise software applications will embed agentic capabilities (Coshow et al., 2024). Technical benchmarks reflect a similar trajectory; performance on OSWorld, a standard benchmark assessing AI agents on real computer tasks, surged from approximately 12% accuracy in early 2024 to 66.3% by the end of 2025, approaching average human performance (Hostinger Research, 2025).

Furthermore, the industrial sector is projected to achieve the highest compound annual growth rate in AI deployment through 2033 (Grand View Research, 2025). These figures underline an important reality: autonomous decision-making components are increasingly incorporated into the design, validation, and governance structures of modern engineering systems.

Key Characteristics and the Concept of ”Agenticness”

No agentic system is simply defined by the size of its underlying language model but by the architecture within which that model operates. Agentic AI is characterized by four core dimensions: multi-step goal-directed planning, persistent contextual memory across interactions, the ability to invoke and coordinate external tools, and behavioral adaptation in response to environmental feedback (Abou-Ali et al., 2025; Acharya et al., 2025).

For example, an agentic architecture can be exemplified by a system that independently queries a sensor network, identifies an anomaly, generates a maintenance recommendation, and logs corrective actions without human prompting at each stage. Unlike traditional automation, which follows fixed logic rules, an agentic system dynamically selects and sequences actions to solve problems whose exact conditions were not fully specified in advance.

Recent academic literature formalizes this capability through the concept of ”agentic-ness”—defined as the degree to which a system adaptably achieves complex goals in dynamic environments with limited direct supervision (Abou-Ali et al., 2025). This concept evaluates autonomy in terms of goal complexity, environmental variability, adaptability, and action scope (Abou-Ali et al., 2025). This formalization offers a rigorous foundation for evaluating autonomy levels during system procurement, deployment, and safety auditing for engineering practitioners.

Sector-specific engineering applications

Agentic architectures are already driving breakthroughs across engineering sectors in real-world applications. Major manufacturers are deploying artificial intelligence (AI)-driven factory operations where autonomous agents coordinate across production lines, anticipate supply chain disruptions, and flag quality anomalies in real time (Chen, 2026). In civil and structural engineering, multi-agent frameworks monitor critical infrastructure assets, processing sensor stream data across bridges, tunnels, and water networks to estimate the remaining useful life and autonomously initiate inspection routines.

The energy and natural resources sector is equally well-suited for agentic deployment. Upstream oil and gas operations require the integration of seismic interpretation, production monitoring, reservoir simulation, and economic evaluation across distributed assets within data-intensive, multidisciplinary workflows (Al-Rawahi et al., 2025). Conventional tools require significant manual intervention during calibration and model updates. Agentic systems streamline these processes by dynamically managing workflows, detecting model misalignments, and adjusting simulation parameters in real time, compressing analysis cycles that previously required lengthy manual iteration (Al-Rawahi et al., 2025).

In engineering design, agentic systems act as proactive collaborators capable of generating design candidates, evaluating performance against structural criteria, identifying failure modes, and iterating toward optimal configurations. This builds upon the materials informatics framework discussed in the January 2024 IJTech editorial, which positioned AI as an accelerator for material discovery (Whulanza & Kusrini, 2024). Agentic AI expands this paradigm by enabling the pursuit of autonomous goals across design solution spaces, thereby reducing the need for explicit human direction at every phase.

Governance, Accountability, and Safety Engineering

However, integrating autonomous decision-making into engineering systems introduces governance challenges that cannot be resolved by standard performance metrics. When an autonomous agent contributes to a structural failure or operational outage, the decision pathway must remain traceable, auditable, and interpretable. Recognizing this, Siemens Energy mandated in early 2026 that all industrial AI decisions in critical infrastructure must be verifiable, auditable, and subject to final human authority (Evolance Market Research, 2026).

The IEEE Standard for Fail-Safe Design of Autonomous and Semi-Autonomous Systems (IEEE, 2024) offers a formal framework to address these requirements, defining criteria for fail-safe behavior, autonomy levels, human approval checkpoints, and emergency shutdowns. Compliance requires embedding classical systems engineering practices—such as functional decomposition, fault-mode analysis, and lifecycle governance—directly into AI system architectures. Evaluating these architectures also requires reproducible and explainable benchmarks tailored to domain-specific tasks (Li & Storhaug, 2026).

Moving forward, agentic AI opens up a rich research area. The critical topics include developing robust control architectures for industrial environments, establishing formal verification methods for non-deterministic agent decisions, and integrating multi-agent systems with digital twins and SCADA networks. Additionally, defining clear boundaries for human-agent collaboration and establishing ethical governance frameworks remain vital research priorities.