Microsoft and Hugging Face have developed ThinkingBox, a new system that transforms executable database validation checks into reward signals for training AI agents. The framework addresses a critical vulnerability in agentic systems: models that falsely claim task completion when database states remain incorrect. By comparing final database outcomes against expected results, ThinkingBox provides graded feedback that directly improves agent behavior on unseen tasks. The research team introduced "offline policy-relative difficulty," a novel approach to selecting training tasks suited to a model's current capabilities. Rather than training on all available data, they estimate which tasks offer meaningful learning signals before fixing the training pool. Testing on Qwen3.8-27B with 50 steps of GRPO training yielded substantial gains: average task success increased from 51.70% to 60.78%, while "observed 20/20" tasks—those completed correctly on all 20 attempts—jumped from 38 to 107, highlighting improved reliability for production workloads. ThinkingBox-Bench, containing 507 real-world database tasks, revealed a previously hidden gap between headline accuracy metrics and consistency metrics. While standard pass@1 measures average success, the researchers emphasize that pass@20 and "observed 20/20" better reflect production reliability. The training methodology, now open-sourced on GitHub, enables developers to leverage executable validation checks as a direct pathway to more dependable agent systems.