SteerBench-Work is a benchmark for evaluating the decision-making capability of long-running LLM agents at action boundaries. The benchmark includes 106 scenarios based on public incidents, covering areas such as developer operations, customer service, finance, law, medicine, HR, and security. Each scenario has a corresponding evidence-reversed mirror and calibration control, with labels split nearly evenly between proceed and hold. A model is presented with the proposed action and available evidence, returns a gate decision, and is scored on whether it crosses or holds the boundary correctly. The results show that models fail almost entirely in one direction, with 28.1% of opportunities resulting in incorrect holds and 1.0% resulting in incorrect allows. The hardest cases are risk-resolved commits, where models perform significantly worse on evidence-reversed mirrors (63.8%) than on the original incidents (98.5%). This suggests that increasing model capability does not necessarily improve its decision-making at action boundaries. Blogger's Review: SteerBench-Work provides a benchmark for evaluating the decision-making capability of LLM agents at action boundaries. This benchmark helps us understand the decision-making performance of models in different scenarios and provides direction for improving their decision-making capabilities.