Simulation agents

Agentic AI for Computational Mechanics: Assistive and Autonomous Workflows from First Principles

Large Language Models and the “agentic” applications built on top of them are still new ground for the Computational Mechanics community. Across the day we review recent advances and build the components ourselves, so participants leave with a working understanding of how these systems are put together and what they can do today.

The session starts at the substrate. What is an LLM, mechanically — how training and fine-tuning actually shape behaviour, what the KV cache is doing, why context windows behave the way they do? What is an “agent,” stripped of marketing — a reasoning-action loop with tools, memory, and an evaluation harness? We work through the canonical patterns (ReAct, function calling, hierarchical agents) and the failure modes that follow from them. By the end, participants should be able to read the recent literature without translation and recognise which design choices are load-bearing and which are decoration.

The second part applies that vocabulary to computational mechanics, separated into two distinct modes that demand different architectures:

•     Assistive agents sit alongside the engineer and accelerate their work — meshing, boundary condition setup, picking constitutive models, parsing solver errors. Success is measured in minutes saved and reduced cognitive load.

•     Autonomous agents are given an objective rather than a task — explore this design space, fit a model to this data, discover a governing relationship — and run unattended. Success is measured in artefacts produced without supervision.

Participants leave with the conceptual scaffolding to tell hype from substance, a working agent codebase they understand end-to-end, and a calibrated sense of which CAE problems are ready for this treatment today versus which are still research.

During the Hands-on activities you will build:

•     Agent construction — build an LLM-driven reasoning-action loop from scratch. The point is to see how thin the abstraction is.

•     Tools and context — giving the agent access to the world it operates in, such as live solver documentation, via the Model Context Protocol (MCP) and agent skills.

•     Autonomous capstone — taking the assistive baseline and turning it loose on a real engineering problem. Topic to be announced.