CFAgentBench: A Reproducible Environment and Benchmark for Autonomous Construction-Finance Agents

CFAgentBench has been introduced as a reproducible environment and benchmark for autonomous construction-finance agents, featuring 1,014 machine-gradeable task specifications across 8 domains. The benchmark utilizes 35 mock applications and employs functional correctness grading with a money-movement guard to ensure tasks requiring human approval are not executed automatically. Initial evaluations show that the strongest agent achieves a pass rate of 0.67 for single attempts, but drops to 0.38 under repeated conditions, highlighting the challenges in deploying reliable construction-finance AI systems.