Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents

Reinforced Agent shifts the evaluation of tool-calling agents from post-hoc checking to in-execution review. A Reviewer Agent checks the primary agent's preliminary tool calls before execution, designed to correct tool selection, parameter accuracy, and scope-awareness errors.

The core idea is separating the Primary Agent from the Reviewer Agent to catch errors within the reasoning loop. However, since the review process can also introduce new errors, the authors proposed a Helpfulness-Harmfulness metric that measures both the benefit and the side effects together.

• Helpfulness: The proportion of the primary agent's errors that the feedback corrected
• Harmfulness: The proportion of originally correct responses that the feedback broke

Evaluated on BFCL (single-turn) and τ2-Bench (stateful multi-turn scenarios), irrelevance detection improved by +5.5%, and multi-turn tasks improved by +7.1%. The choice of reviewer model also mattered: o3-mini showed a benefit-to-harm ratio of 3:1, while GPT-4o only reached 2.1:1. GEPA-based automatic prompt optimization added a further +1.5~2.8%, demonstrating that the reviewer model and prompt can be improved without retraining the primary agent.