Apple Paper Tests Real-Time Review for Tool-Calling Agents
Apple researchers have published a new paper proposing that tool-calling agents can perform better when a dedicated reviewer checks their actions before execution, not after. The approach, outlined in "Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents," delivered measurable gains on two established benchmarks, though the team also warns that the reviewer itself can degrade otherwise correct outputs.
How Inference-Time Feedback Works
Most evaluation methods for tool-calling agents operate after the fact. They catch errors in tool selection, parameter formatting, or scope recognition, but any fixes require prompt adjustments or full retraining. The Apple team, led by researchers Anh Ta, Junjie Zhu, and Shahin Shayandeh, takes a different path by embedding a secondary review agent directly into the execution loop.
During inference, the reviewer evaluates provisional tool calls before they run. This shifts the process from reactive error cleanup to proactive course correction. The architecture separates the primary execution agent from the review agent, giving each a distinct role.
To measure whether the reviewer actually helps, the paper introduces a pair of metrics it calls Helpfulness-Harmfulness. Helpfulness tracks how many base-agent errors the reviewer successfully fixes. Harmfulness captures the opposite: how often feedback turns a correct response into an incorrect one. Together, the two scores reveal whether a given reviewer configuration provides net value or creates more problems than it solves.
Benchmark Results and Model Tradeoffs
Testing on BFCL, a single-turn benchmark, the system improved irrelevance detection by 5.5%. On τ2-Bench, which covers multi-turn stateful scenarios, performance rose by 7.1%. The gains depended heavily on which model served as the reviewer.
The reasoning model o3-mini reached a 3-to-1 benefit-to-risk ratio, meaning it fixed three errors for every correct answer it broke. GPT-4o managed a lower 2.1-to-1 ratio. The researchers also applied an automated prompt optimization technique called GEPA, which added another 1.5% to 2.8% on top of those results.
What This Means for Agent Development
The core takeaway is practical: separating execution from review lets teams upgrade the reviewer through model swaps and prompt tuning without retraining the base agent. That modularity could lower the cost of iterating on agent reliability.
Still, the Helpfulness-Harmfulness framework highlights a real tension. Every reviewer introduces some risk of corrupting good outputs. The paper argues that no prior work has systematically quantified this tradeoff, making the new metrics a useful diagnostic tool for anyone building multi-agent systems.
The paper was accepted at the Fifth Workshop on Natural Language Generation, Evaluation, and Metrics at ACL 2026. The full research is available on Apple's Machine Learning Research page.