New benchmark evaluates AI agents on complex real-world architecture and construction tasks using multimodal data

AEC-Bench tests agentic systems on Architecture, Engineering, and Construction domain tasks including drawing understanding and cross-sheet reasoning

Researchers identified tool usage and harness design techniques that consistently improve performance across foundation models like Claude Code and Codex

Full benchmark dataset, agent harness, and evaluation code released open-source under Apache 2 license on GitHub for reproducibility