Open‑ended tool agents can achieve the same objective via different tool paths, and their quality depends on the interaction trace, environment state and final artifacts, making an executable ground‑truth reward rare. The main challenge is to build a reward from existing experience that reveals the current policy’s failures while staying stationary during optimization. Reward Engine works like a self‑motivated student: first determine what success requires, then optimize toward it. From the task description and initial trajectories it constructs a RewardPack that contains testcases for required correctness and rubrics for graded quality, each check returning a judgment with evidence. The frozen RewardPack guides the agent to revise and select its outputs and can also serve as a reinforcement‑learning reward. On WorkBuddy Bench Office and Claw‑Eval the constructed reward largely agrees with benchmark graders, not reversing their ordering on 82.3% and 80.5% of candidate pairs. When used to revise executions it raises three‑run consistency $\mathrm{Pass}^{3}$ from 45.0% to 54.0% and from 78.4% to 83.7%.
Review