Sim Agent: Building an AI Assistant for a Scientific Simulator

Building a scientific simulator to quantify light’s behavior for quantum-sensing is hard. Operating this type of software as a user can be even harder! At Eigenlabs, we developed an optical quantum-sensing simulator that is usable without requiring a PhD in quantum physics. To achieve this, we developed "Sim Agent:" a dual-workflow system that pairs a standard web user interface with an embedded LLM agent. The agent understands all the capabilities of the software and can configure, run, analyze, document, and discuss complex physics simulations.

This talk breaks down the architecture and lessons learned from building our domain-specific scientific assistant. We will explore key technical strategies, including crafting system prompts for strict behavioral boundaries, agent tools for deterministic math to eliminate hallucinations in precision calculations, and designing robust agent loops. We will also cover an innovative self-monitoring capability which enables the agent to automatically log edge-case failures or self-submit feature requests with full simulation parameters for instant developer reproducibility.

Attendees will gain a practical blueprint for embedding AI assistants into domain-specific software. You will learn how to: balance agent autonomy with deterministic precision, handle strict domain constraints, and implement automated feedback loops that turn agent failures into new features.