EDGEGEN: Improving Tool-Calling Agents Beyond Happy Paths with Synthetic Edge Case Generation

This paper proposes EdgeGen, a synthetic task generation framework that extracts compliance rules from agent specifications to produce database-grounded edge-case tasks designed to violate these rules. Operating as a fully automated closed-loop system without human annotation, EdgeGen facilitates agent improvement through fine-tuning and harness optimization. Fine-tuning models on data generated by EdgeGen yields consistent mean progress improvements ranging from 2 percent to 42 percent on the tau2bench airline domain. Additionally, for harness optimization, the method achieves mean progress improvements of 10 percent and 30 percent over human-curated and base harnesses respectively on the Gemma-4-e4b model.

Finetuning on data generated by EdgeGen yields a consistent mean progress improvement of 2 percent to 42 percent on tau2bench airline domain, while other baseline methods show degradation for some models.

For harness optimization, our method shows a mean progress improvement of 10 percent and 30 percent over the human-curated and base harnesses, respectively, for the Gemma-4-e4b model.