---

To learn about the latest thinking about Agentic AI within mechanical engineering—where independent AI modules interact to solve complex problems—I consulted the popular AI assistant Grok. Here is what came up:

An Agentic AI in mechanical engineering is an autonomous software system that can independently pursue complex, open-ended engineering goals with minimal human intervention by repeatedly perceiving the current state, reasoning about it, planning actions, executing those actions (usually via tools or APIs), and learning from the outcomes. In mechanical engineering practice, this typically manifests as one or more of the following concrete capabilities:

1. Design Autonomy
   • Takes a high-level requirement (e.g., “design a 50 kW gearbox for an EV with < 0.5% efficiency loss at 12,000 rpm, under 12 kg, cost < $180 in 100k volume”).
   • Iteratively generates CAD/solid models, runs topology optimization, selects materials, performs FEA/CFD, checks manufacturability (DFM), iterates, and outputs production-ready STEP files, drawings, and a BOM — all without human-in-the-loop after the initial prompt.
2. Simulation & Validation Agency
   • Sets up its own multi-physics simulations (Ansys, Abaqus, STAR-CCM+, OpenFOAM, etc.), writes input decks or Python scripts, runs parameter sweeps or formal design optimization (Bayesian, genetic, gradient-based), interprets results, and decides whether the design meets specs or needs another loop.
3. Physical Test Autonomy
   • Generates G-code or robot paths, instructs collaborative robots or test rigs via API to build and test prototypes, reads sensor data (torque, vibration, thermography, strain gauges), compares to simulation, and closes the digital–physical loop.
4. Supply-Chain & Manufacturing Agency
   • Queries supplier APIs for real-time pricing/lead times, runs should-cost models, negotiates (in some advanced cases), selects processes (additive vs. CNC vs. casting), generates CAM toolpaths, and pushes work orders to MES systems.
5. Lifecycle & Maintenance Agency
   • In service, an agentic system can ingest fleet telemetry, diagnose failures, propose redesigned parts, validate them virtually, and push OTA updates to reconfigurable hardware (e.g., metamaterial lattices or adjustable mechanisms).

Upon reading this, I thought it to be an extremely ambitious aspirational statement, and not more than that. Then Grok added:

Based on expert forecasts, scaling laws, and industry signals, we’re 2-5 years from meaningful realization of agentic AI handling integrated workflows like full product lifecycle management in mechanical engineering, with human oversight reduced to 20-30%. This aligns with broader AI trends, where timelines have compressed dramatically.

In my opinion, this forecast is completely unrealistic for the following reason: In a multi-agent system, an error made by one agent will propagate to others, leading to poor or even nonsensical outcomes. Therefore, tight control of model form and approximation errors is essential. While the theoretical understanding of how to control these errors does exist, the vast majority of software tools used in industry, including those mentioned under the heading Simulation & Validation Agency, lack the requisite capabilities and are thus unsuited for incorporation into agentic AI systems.

Any effort to develop an agentic system with some or most of the functionalities listed above should, first and foremost, focus on controlling errors in numerical simulation through the application of verification, validation, and uncertainty quantification procedures.

Building an agentic system on top of legacy finite element modeling practices is like building a skyscraper on quicksand. Any team serious about reaching these capability levels must treat the soundness of numerical simulation as the highest priority, not an afterthought.

---

Related Blogs:

• A Memo from the 5th Century BC
• Obstacles to Progress
• Why Finite Element Modeling is Not Numerical Simulation?
• XAI Will Force Clear Thinking About the Nature of Mathematical Models
• The Story of the P-version in a Nutshell
• Why Worry About Singularities?
• Questions About Singularities
• A Low-Hanging Fruit: Smart Engineering Simulation Applications
• The Demarcation Problem in the Engineering Sciences
• Model Development in the Engineering Sciences
• Certification by Analysis (CbA) – Are We There Yet?
• Not All Models Are Wrong
• Digital Twins
• Digital Transformation
• Simulation Governance
• Variational Crimes
• The Kuhn Cycle in the Engineering Sciences
• Finite Element Libraries: Mixing the “What” with the “How”
• A Critique of the World Wide Failure Exercise
• Meshless Methods
• Isogeometric Analysis (IGA)
• Chaos in the Brickyard Revisited
• Why Is Solution Verification Necessary?
• Variational Crimes and Refloating the Costa Concordia
• Lessons From a Failed Model Development Project
• Where Do You Get the Courage to Sign the Blueprint?