A new comparative study examines how large language models can be deployed to automate modifications to building energy simulation models, with direct implications for utilities managing efficiency programs and designers undertaking retrofits under energy codes.
Researchers evaluated two LLM agent architectures tasked with editing EnergyPlus models—the industry-standard building physics simulation platform—to comply with Canada's National Energy Code for Buildings (NECB). The tool-augmented approach provides the LLM with pre-built, domain-specific functions that encapsulate HVAC engineering logic and safe model manipulation routines. The autonomous approach grants the LLM direct access to primitive operations: creating, deleting, and modifying individual model objects and their properties.
Both systems were implemented as multi-agent workflows using the Model Context Protocol, integrating retrieval-augmented access to NECB standards, a simulation execution agent, and a coordinator. Testing spanned seven reference building types—apartments, restaurants, schools, warehouses, and offices—across three modification tasks of increasing complexity.
For simple tasks (modifying window U-values), both approaches achieved 100% success. For medium-complexity tasks (reducing lighting power), both succeeded across all buildings, though autonomous execution consumed more tokens and processing time. Performance diverged sharply on complex tasks: converting HVAC systems from packaged terminal (PTAC) or direct-expansion (DX) units to variable refrigerant flow (VRF) systems.
The tool-augmented approach succeeded in all seven buildings, producing valid, NECB-compliant models ready for annual energy simulation. The autonomous approach failed completely: apartment conversions stalled during planning, and packaged-DX conversions produced structurally invalid models that could not be simulated.
These results suggest a pragmatic hybrid strategy: autonomous agents for simple, low-interdependency edits; validated domain-specific tools for complex, compliance-critical transformations where coordinated changes and code adherence are non-negotiable.
